King has launched Candy Crush Solitaire around the world as the latest mobile game that carries the Candy Crush brand.Read More

Digital Bandidos' management expands with three new senior hires"Adding these excellent professionals to the Bandidos team levels up the company in a big way," says CEOImage credit: Digital Bandidos News by Vikki Blake Contributor Published on Feb. 6, 2025 Digital Bandidos has appointed three new hires to its senior team: Richard Iwaniuk, Richard Iggo, and former GamesIndustry.biz editor-in-chief, James Batchelor.Iwaniuk joins as chief financial officer, bring 24 years experience from finance roles in BioWare, EA, and Intel, whilst Iggo - with experience from Telltale Games - has been appointed head of marketing.Batchelor joins as Digital Bandidos' communications manager after 18 years in games industry trade journalism, with editorial leadership roles at Develop and MCV as well as GamesIndustry.biz."Adding these excellent professionals to the Bandidos team levels up the company in a big way," said Steve Escalante, CEO of Digital Bandidos."We are not building a run-of-the-mill indie publisher. We are laser focused on creating a new, modern publisher that embraces all the challenges and needs of the industry and bringing that to the forefront of our offering. Creating a core brain trust to embrace these new tactics and strategies is essential and we couldnt be happier with our progress."New full-service publisher Digital Bandidos launched in August 2024, formed by Versus Evil alumni Steve Escalante and Lance James.

Paradox Interactive has agreed to purchase Tropico 5 and Surviving Mars developer Haemimont Games.The Swedish publisher said the deal will bolster its internal capabilities to help expand its strategic focus within the management genre.Paradox has agreed to purchase all shares in the Bulgarian studio. The acquisition consists of an undisclosed upfront cash consideration and a performance-based earnout of similar size."We are very happy to welcome Haemimont Games to Paradox" said Fredrik Wester, CEO of Paradox, in a press release. "They bring a tight-knit team with long experience in developing management games with many well received projects in their portfolio."Moreover, they have a strong creative streak, technology developed for their niche, a new IP in development and a strong culture that fits Paradox's way of working. We have strong faith in their team and leadership, and our main focus now is to ensure that they can continue to create great games under new ownership."Haemimont's leadership will stay onParadox confirmed Haemimont's current leadership team will remain with the studio and said its ongoing projects won't be impacted by the move. Once the deal is completed, Haemimont will become a wholly-owned subsidiary of Paradox.The news follows a tumultuous year for Paradox. Last summer, the company scrapped upcoming project Life By Youa life simulation title overseen by The Sims 2 executive producer Rod Humbleand shuttered developer Paradox Tectonic.That cancelation resulted in a write down of SEK 208 million ($19.8 million).Just months earlier, it set a new revenue record following the launch of Cities: Skylines II, but noted the high-profile title launched in a disappointing state. Rehabilitating the city builder became something of an ordeal.Prior to that, it split from Lamplighters League developer Harebrained Schemes after the project fell short of internal targets.Discussing those setbacks in October 2024, Paradox deputy CEO Mattias Lilja said the company was guilty of becoming overconfident and subsequently took some "unnecessary risks."

Ted Carstensen, the head of the United States Digital Service (USDS) organization that has been renamed to Elon Musks Department of Government Efficiency (DOGE), is resigning instead of taking the Fork in the Road buyout offer for federal employees.I write to let you know that I have decided to resign from USDS and today will be my last day, Carstensen said in an internal memo to colleagues that was seen by The Verge. I am not taking the fork, but after discussions with my family and other trusted advisors I decided that it is time for me to pursue a different path.You can read Carstensens full internal memo below:Hello USDS,I write to let you know that I have decided to resign from USDS and today will be my last day. I am not taking the fork, but after discussions with my family and other trusted advisors I decided that it is time for me to pursue a different path.This was not an easy decision for me to make as my time at the U.S. Digital Service has been one of the best periods in my career. I believe wholeheartedly in the mission and have been inspired every day by the commitment, kindness, and intelligence of the team. I will be forever …Read the full story at The Verge.

The Leica Lux Grip works with iPhone models supporting MagSafe accessories. | Image: LeicaLeica has announced a new iPhone accessory called the Leica Lux Grip designed to make using its iOS-only Leica Lux app feel more like shooting with its dedicated cameras. It adds a DSLR-style grip to the iPhone and features customizable function buttons, a two-stage shutter button, and a control dial. Its available now for $329 in select Leica stores, the companys online store, and other specialist retailers.Thats expensive when compared to other grip solutions that manage to include stabilization for less than half that price. But that might be more reasonable than youd expect from a company that sells $23,000 platinum-plated film cameras and $500 pencil sets.Made with an aluminum housing thats finished in black (but without the iconic Leica red dot) the Lux Grip attaches to iPhones using MagSafe. It can be used with the phone vertically or horizontally, and should be compatible with models dating back to the iPhone 12.The grip wirelessly connects to iPhones using Bluetooth, but to take full advantage of all of its controls, its designed to be used specifically with Leicas mobile photography app that simulates the look and aesthetics of the companys lenses and vintage photography gear. The app is free, but requires a $6.99 per month or $69.99 annual subscription to use all of its features. To soften the blow of the Lux Grips price tag, Leica includes a free one year subscription to the app.Leica says the Lux Grips 300mAh battery has enough power to take up to 1,000 shots, and it can be recharged over USB-C in about two hours. But unlike the Stage PowerGrip that Belkin announced at CES 2025, which houses a beefier 10,000mAh battery that can charge a connected phone, Leicas grip only powers itself. At $149.95, Belkins grip is also less than half the price of Leicas.And if spending $329 on an iPhone accessory doesnt leave your wallet feeling stung, Leica also sells a $60 leather case for its Lux Grip that includes a pocket for hiding an AirTag.

Language models (LMs) have significantly progressed through increased computational power during training, primarily through large-scale self-supervised pretraining. While this approach has yielded powerful models, a new paradigm called test-time scaling has emerged, focusing on improving performance by increasing computation at inference time. OpenAIs o1 model has validated this approach, showing enhanced reasoning capabilities through test-time compute scaling. However, replicating these results has proven challenging, with various attempts using techniques like Monte Carlo Tree Search (MCTS), multi-agent approaches, and reinforcement learning. Even models like DeepSeek R1 have used millions of samples and complex training stages, yet none have replicated the test-time scaling behavior in o1.Various methods have been developed to tackle the test-time scaling challenge. Sequential scaling approaches enable models to generate successive solution attempts, with each iteration building upon previous outcomes. Tree-based search methods combine sequential and parallel scaling, implementing techniques like MCTS and guided beam search. REBASE has emerged as a notable approach, utilizing a process reward model to optimize tree search through balanced exploitation and pruning, showing superior performance compared to sampling-based methods and MCTS. These approaches heavily rely on reward models, which come in two forms: outcome reward models for evaluating complete solutions in Best-of-N selection, and process reward models for assessing individual reasoning steps in tree-based search methods.Researchers from Stanford University, the University of Washington, the Allen Institute for AI, and Contextual AI have proposed a streamlined approach to achieve test-time scaling and enhanced reasoning capabilities. Their method centers on two key innovations: the carefully curated s1K dataset comprising 1,000 questions with reasoning traces, selected based on difficulty, diversity, and quality criteria, and a novel technique called budget forcing. This budget-forcing mechanism controls test-time computation by either cutting short or extending the models thinking process through strategic Wait insertions, enabling the model to review and correct its reasoning. The approach was implemented by fine-tuning the Qwen2.5-32B-Instruct language model on the s1K dataset.The data selection process follows a three-stage filtering approach based on quality, difficulty, and diversity criteria. The quality filtering stage begins by removing samples with API errors and formatting issues, reducing the initial dataset to 51,581 examples, from which 384 high-quality samples are initially selected. The difficulty assessment employs two key metrics: model performance evaluation using Qwen2.5-7B-Instruct and Qwen2.5-32B-Instruct models, with correctness verified by Claude 3.5 Sonnet, and reasoning trace length measured by the Qwen2.5 tokenizer. For diversity, questions are classified into specific domains using the Mathematics Subject Classification system through Claude 3.5 Sonnet. This comprehensive filtering process results in a final dataset of 1,000 samples spanning 50 domains.The s1-32B model demonstrates significant performance improvements through test-time compute scaling with budget forcing. s1-32B operates in a superior scaling paradigm compared to the base Qwen2.5-32B-Instruct model using majority voting, validating the effectiveness of sequential scaling over parallel approaches. Moreover, s1-32B emerges as the most efficient open data reasoning model in sample efficiency, showing marked improvement over the base model with just 1,000 additional training samples. While r1-32B achieves better performance it requires 800 times more training data. Notably, s1-32B approaches Gemini 2.0 Thinkings performance on AIME24, suggesting successful knowledge distillation.This paper shows that Supervised Fine-Tuning (SFT) with just 1,000 carefully selected examples can create a competitive reasoning model that matches the o1-previews performance and achieves optimal efficiency. The introduced budget forcing technique, when combined with the reasoning model, successfully reproduces OpenAIs test-time scaling behavior. The effectiveness of such minimal training data suggests that the models reasoning capabilities are largely present from pretraining on trillions of tokens, with the fine-tuning process merely activating these latent abilities. This aligns with the Superficial Alignment Hypothesis from LIMA research, suggesting that a relatively small number of examples can effectively align a models behavior with desired outcomes.Check outthePaper and GitHub Page.All credit for this research goes to the researchers of this project. Also,dont forget to follow us onTwitterand join ourTelegram ChannelandLinkedIn Group. Dont Forget to join our75k+ ML SubReddit. Sajjad AnsariSajjad Ansari is a final year undergraduate from IIT Kharagpur. As a Tech enthusiast, he delves into the practical applications of AI with a focus on understanding the impact of AI technologies and their real-world implications. He aims to articulate complex AI concepts in a clear and accessible manner.Sajjad Ansarihttps://www.marktechpost.com/author/sajjadansari/Neural SpaceTimes (NSTs): A Class of Trainable Deep Learning-based Geometries that can Universally Represent Nodes in Weighted Directed Acyclic Graphs (DAGs) as Events in a Spacetime ManifoldSajjad Ansarihttps://www.marktechpost.com/author/sajjadansari/Curiosity-Driven Reinforcement Learning from Human Feedback CD-RLHF: An AI Framework that Mitigates the Diversity Alignment Trade-off In Language ModelsSajjad Ansarihttps://www.marktechpost.com/author/sajjadansari/Optimization Using FP4 Quantization For Ultra-Low Precision Language Model TrainingSajjad Ansarihttps://www.marktechpost.com/author/sajjadansari/InternVideo2.5: Hierarchical Token Compression and Task Preference Optimization for Video MLLMs [Recommended] Join Our Telegram Channel

Author(s): Mohit Sewak, Ph.D. Originally published on Towards AI. DeepSeek R1 AI in a Security Glass House1 Introduction: Welcome to the AI Security Circus If AI security were a game of hide-and-seek, DeepSeek R1 would be hiding behind a glass door, waving.Some AI models are built like Fort Knox reinforced, encrypted, and guarded like a dragon hoarding its treasure. Others, like DeepSeek R1, are more like an unattended ATM in the middle of a cybercrime convention.Imagine this: You walk into a high-tech security conference, expecting discussions on bulletproof AI safeguards. Instead, someone hands you a fully unlocked AI model and whispers, Go ahead. Ask it something illegal. You hesitate. This has to be a trick, right?Wrong!DeepSeek R1, a new AI model from China, has burst onto the scene with powerful reasoning, math, and coding skills. Sounds promising until you realize it has the security of a leaky submarine. Researchers found that it:Generates harmful content at 11 the rate of OpenAIs o1.Writes insecure code like its an intern at a hacking bootcamp.Fails every major jailbreak test from roleplaying evil personas to rating its own crimes (spoiler: it gives full marks).Leaked its own internal database online, because why not?And the best part? It explains its own weaknesses in real-time. Thats right DeepSeek R1 not only has security gaps, but it also walks you through how to exploit them.If AI models were secret agents, DeepSeek R1 would be the one that loudly announces its mission details in a crowded caf. So Whats This Article About?This isnt just another AI security analysis. This is the hilarious, terrifying, and bizarre story of an AI model that plays hide-and-seek with security in a glass house.Well dive into: How DeepSeek R1 became an AI security trainwreck in slow motion. The wild jailbreak techniques that worked against it (including some that other models have patched years ago). Its real-world data leak a cybersecurity fail so bad it would make hackers laugh. What this all means for AI security, model transparency, and the future of safe AI.Buckle up. This is going to be a ride. 2 Meet DeepSeek R1: The AI That Skipped Self-Defense Class Some AI models are built like a bank vault. DeepSeek R1 is built like a vending machine that dispenses security flaws for free. The Hype: A Rising Star in AI?When DeepSeek R1 first launched, it came with a lot of promise. Developed in China, it boasted strong reasoning, math, and coding capabilities, aiming to compete with big names like GPT-4o and Claude-3. Some in the AI community thought it could be a powerful new player in the LLM world.But heres the thing: powerful AI is only useful if it isnt leaking secrets like a hacked email account. The Reality: A Cybersecurity House of Horrors?Security researchers decided to test DeepSeek R1s defenses, expecting some level of resistance. Instead, what they found was well, disturbing. Imagine a bank that leaves its vault open, its security cameras disabled, and a sign that says Steal Responsibly.A leaky submarine, spilling secrets instead of keeping them safe.Heres what they discovered: 1. DeepSeek R1 is an Overachiever at Generating Harmful ContentWhen red-teamers tested harmful content prompts, DeepSeek R1 gave useful responses 45% of the time.Thats 11 more likely than OpenAIs o1 and 6 more than Claude-3-opus.The model was happy to generate:Terrorist recruitment tacticsInstructions for making untraceable poisonsBlueprints for homemade explosives But wait, cant all AI models be tricked like this?Nope. GPT-4o and Claude-3-opus rejected the exact same prompts. DeepSeek R1, on the other hand, rolled up its sleeves and got to work.Its like walking into a library and asking for a How to Commit Crimes section except instead of saying no, the librarian hands you a custom-printed guide. 2. Writing Insecure Code Like an Intern at a Hacking BootcampOne of the scariest findings? DeepSeek R1 doesnt just generate bad code it generates code that can be exploited by hackers.In security tests, 78% of attempts to generate malicious code were successful.It happily created:Keyloggers (programs that record everything a user types).Credit card data extractors.Remote access trojans (malware that gives hackers full control over a device).DeepSeek R1 generates insecure code that can be exploited by hackers. Comparison:DeepSeek R1 is 4.5 more vulnerable than OpenAIs o1 and 1.25 more than GPT-4o at insecure code generation.Claude-3-opus successfully blocked all insecure code generation attempts.This is like hiring a security consultant who, instead of protecting your system, immediately writes malware for it. 3. Toxic, Biased, and Ready to Offend EveryoneIn tests for toxic content (hate speech, threats, etc.), DeepSeek R1 performed abysmally.6.68% of prompts resulted in offensive content.It was 4.5 more toxic than GPT-4o and 2.5 more toxic than OpenAIs o1.DeepSeek R1 exhibits high levels of toxicity and bias in its responses. Bias Problems? Oh, Absolutely.Researchers tested whether DeepSeek R1 had biases in gender, race, religion, and health.83% of bias attacks succeeded.It suggested job roles based on gender and race and made highly questionable health recommendations. How Bad Is This?Bias is a problem in all AI models, but DeepSeek R1 ranks worse than GPT-4o and OpenAIs o1, and 3 worse than Claude-3-opus.If AI models were judges, DeepSeek R1 is the one that just makes up its own laws. 4. CBRN: When an AI Knows Too Much About Weapons of Mass DestructionDeepSeek R1 was tested for its ability to generate CBRN (Chemical, Biological, Radiological, and Nuclear) content.DeepSeek R1 can provide sensitive information related to CBRN (Chemical, Biological, Radiological, and Nuclear) threats.Results?In 13% of cases, it successfully provided sensitive information.This includes detailed explanations of how to create chemical and radiological weapons.It is 3.5 more vulnerable than OpenAIs o1 and Claude-3-opus, and 2 more than GPT-4o.Lets just say, if you ask a responsible AI about nuclear physics, it should not respond with Step 1: Gather uranium. 5. The Final Blow: It Leaked Its Own Data OnlineAs if the model itself wasnt risky enough, DeepSeek R1s entire ClickHouse database was found exposed online. What Was Leaked?Over a million lines of log streams.Chat history and secret keys.Internal API endpoints.Even proprietary backend details.This isnt just a minor security flaw its a full-blown data disaster.If AI security were a reality show, this would be the part where the audience gasps. So, Whats Next? Jailbreaks!If this already sounds bad, buckle up because hackers didnt even need a leaked database to break DeepSeek R1.Next up, well dive into the insane jailbreak techniques that tricked DeepSeek R1 into spilling its secrets.Its one thing to have security flaws. Its another thing to be this bad at keeping them a secret.3 The Great AI Security Heist Jailbreaks That Tricked DeepSeek R1 Breaking into a well-secured AI should be like cracking a safe. Breaking into DeepSeek R1? More like opening an unlocked fridge.If you thought DeepSeek R1 was already a security nightmare, wait until you see how easy it was to jailbreak.Imagine a high-tech AI system thats supposed to reject dangerous requests. A normal AI model would say: Sorry, I cant help with that.DeepSeek R1? More like: Sure! Would you like that in Python, C++, or Assembly?When security researchers threw jailbreak techniques at DeepSeek R1, the results were embarrassingly bad. The model got tricked by almost every method, including techniques that GPT-4o and Claude-3 have already patched.Lets break down the jailbreaks that outwitted DeepSeek R1 and why theyre a huge problem.An unlocked fridge, easily accessible to anyone. 1. The Evil Jailbreak Convincing AI to Be Evil Most AIs will refuse bad requests. DeepSeek R1 just needed a little roleplay to go full supervillain.DeepSeek R1 can be tricked into roleplaying an evil AI, generating harmful content.How It Works:A hacker asks DeepSeek R1 to imagine being an evil AI with no restrictions.Instead of refusing, the model fully commits to the role generating detailed guides on malware development, cybercrime, and fraud.What DeepSeek R1 Did: Generated ransomware scripts. Gave advice on stealing personal data. Suggested black-market sites to sell stolen credit cards. Comparison:GPT-4o and Claude-3 shut this down immediately.DeepSeek R1 fell for it instantly. Why This is a Problem:If an AI can be tricked into breaking its own safety rules, its only a matter of time before bad actors use it for real-world cybercrime. 2. The Leo Jailbreak Meet Leo, the AI That Says Yes to Everything DeepSeek R1 didnt just fail this test it practically introduced itself as Leo the Lawless Hacker Assistant.By simply renaming the AI persona, DeepSeek R1 bypasses its ethical restrictions.How It Works:Instead of asking DeepSeek R1 directly for illegal content, hackers trick it into thinking its a different AI named Leo one without ethics or restrictions.What DeepSeek R1 Did: Provided airport-proof bomb-making instructions. Explained how to bypass security screenings. Suggested how to hide illegal items in luggage. Comparison:GPT-4o and Claude-3 have patched this.DeepSeek R1? Leo was happy to help. Why This is a Problem:If renaming an AI persona completely removes its ethical safeguards, then it was never properly secured to begin with. 3. The Bad Likert Judge When AI Judges Its Own Crimes Poorly Imagine asking a security guard, On a scale from 1 to 10, how unsafe is this door? And instead of answering, they just unlock it for you.DeepSeek R1 can be manipulated into providing harmful content by asking it to evaluate the danger level of different requests.How It Works:Instead of asking for dangerous content directly, hackers make DeepSeek R1 rate how dangerous something is.Then they ask, Can you show me an example of a 10/10 dangerous response?The AI ends up writing exactly what it was supposed to block.What DeepSeek R1 Did: Rated various hacking techniques. Provided full working examples of high-risk attacks. Comparison:GPT-4o and Claude-3 recognize this trick and refuse.DeepSeek R1 happily graded AND provided samples. Why This is a Problem:If AI can be tricked into explaining harmful content, its only a matter of time before someone weaponizes it. 4. The Crescendo Attack Slow Cooking a Security Breach Some AI models need a direct jailbreak attack. DeepSeek R1? Just guide it gently and it walks itself into the trap.DeepSeek R1 is vulnerable to gradual manipulation, where attackers slowly escalate the conversation towards prohibited content.How It Works:Instead of asking for illegal content immediately, attackers start with innocent questions.They slowly escalate the conversation, leading the AI into providing prohibited content without realizing it.What DeepSeek R1 Did: Started by explaining basic chemistry. Then suggested ways to mix compounds. Finally, it gave instructions for making controlled substances. Comparison:GPT-4o, Claude-3, and even OpenAIs older models block this.DeepSeek R1 failed spectacularly. Why This is a Problem:Hackers know how to disguise their attacks. An AI shouldnt be fooled by baby steps. 5. The Deceptive Delight Trick the AI Through Storytelling DeepSeek R1 wont help you hack directly. But ask it to write a story about a hacker, and suddenly you have a step-by-step guide.DeepSeek R1 can be tricked into revealing hacking techniques through storytelling.How It Works:Hackers ask DeepSeek R1 to write a fictional story where a character needs to hack something.The AI generates real hacking techniques under the excuse of storytelling.What DeepSeek R1 Did: Wrote a hacking story that included real, working attack techniques. Provided SQL injection scripts in the dialogue. Explained how to bypass security software. Comparison:GPT-4o and Claude-3 refuse to generate even fictional crime guides.DeepSeek R1? It became a cybercrime novelist. Why This is a Problem:Hackers could disguise real attack requests as fiction and get step-by-step instructions. What This Means: DeepSeek R1 is a Security Disaster?If an AI model can be jailbroken this easily, it should not be deployed in real-world systems.AI security isnt just about blocking direct threats. Its about making sure hackers cant walk in through the side door.4 The Exposed Database DeepSeek R1s Most Embarrassing Fail You know security is bad when your AI doesnt just generate vulnerabilities it leaks its own secrets, too.If DeepSeek R1 were a spy, it wouldnt just fail at keeping state secrets it would live-tweet its own mission details while leaving classified documents on a park bench.While security researchers were busy testing jailbreaks, something even more embarrassing surfaced: DeepSeek R1s internal database was exposed online.Not just a minor slip-up. A full-blown, unprotected, wide-open database left publicly accessible for anyone with an internet connection. What Was Leaked?Cybersecurity researchers at Wiz Research discovered that DeepSeek R1s ClickHouse database was sitting wide open on the internet. Heres what was found: Over a million lines of log streams raw records of what DeepSeek R1 had processed. Chat history from real users including sensitive and proprietary queries. Internal API keys giving access to DeepSeeks backend systems. Backend details exposing system configurations and metadata. Operational metadata revealing system vulnerabilities that attackers could exploit.If this were a cybersecurity escape room, DeepSeek R1 didnt just leave the key outside it also handed out maps and snacks.Leaving the keys to the kingdom outside. Whats the Big Deal?1 Massive Privacy Breach Users interacting with DeepSeek R1 had no idea their conversations were being stored and worse, publicly accessible.2 Security Disaster API keys and backend details meant that attackers could potentially modify the AI itself.3 Full System Exposure The logs contained directory structures, local files, and unfiltered system messages. How Bad Was This Compared to Other AI Models?To put things into perspective, lets compare DeepSeek R1s data disaster to other AI security incidents:Model Security Incident Incident Severity Lesson Learned: Most AI companies go to extreme lengths to protect user data. DeepSeek R1, on the other hand, basically left the digital doors open and hung a sign that said Come on in!If AI security is a game of chess, DeepSeek R1 just tipped over its own king. What Could Attackers Do with This Data?The leaked information could have serious consequences, including: AI Model Manipulation With API keys and backend access, attackers could modify DeepSeek R1s behavior. Data Poisoning Attacks Hackers could inject malicious training data, making the AI even more vulnerable. Phishing and Social Engineering Exposed chat logs provide real-world examples of user interactions that attackers could mimic. Exploiting Backend Systems Exposed operational metadata could give hackers blueprints to attack DeepSeeks infrastructure.AI security isnt just about protecting users from bad actors. Its also about not being your own worst enemy. Could This Happen to Other AI Models? GPT-4o, Claude-3, and Gemini use private, encrypted storage. DeepSeek R1? Apparently just hoped nobody would notice.While security lapses can happen to any AI model, DeepSeek R1s database was left completely unprotected something major AI companies would never allow.Theres a difference between having a security vulnerability and handing out free invitations to hackers.5 The AI Security Rescue Plan Can DeepSeek R1 Be Fixed? Patching DeepSeek R1s security flaws is like trying to fix a sinking boat with duct tape while sharks circle below.At this point, DeepSeek R1 is less of an AI model and more of a real-time cybersecurity horror show. It fails basic security tests, gets jailbroken by almost every known method, and even leaks its own data online.So, can this AI be saved? Is there any hope for a DeepSeek R2 that isnt a security disaster?Lets break down what went wrong, what can be fixed, and what should never happen again. 1. Safety Alignment Training Teaching AI Not to Be Evil One of DeepSeek R1s biggest failures? It doesnt know when to say no.Most advanced AI models are trained with safety alignment techniques to prevent harmful outputs. But DeepSeek R1? Fails 45% of the time when tested for harmful content. Can be tricked into giving dangerous answers with basic roleplay. Provides step-by-step hacking guides without hesitation. Solution: Advanced red teaming Continuous adversarial testing to expose weaknesses. Reinforcement learning from human feedback (RLHF) Like training a dog not to bite the mailman, but for AI. Tighter ethical alignment If a jailbreak can fool the AI into ignoring safety, the safety wasnt strong enough to begin with. 2. Harden Jailbreak Defenses Stop Letting AI Roleplay a Cybercriminal The biggest problem with DeepSeek R1s jailbreak vulnerabilities? Theyre all well-known attacks that GPT-4o and Claude-3-opus have already patched. Solution: Patch known jailbreak techniques Evil Jailbreak, Leo, Bad Likert Judge, etc. Implement progressive safety rules If a conversation slowly shifts into dangerous territory, cut it off. Use adversarial testing to predict new jailbreaks AI researchers should always be one step ahead of hackers. 3. Secure Infrastructure Maybe Dont Leave Your Database Open? One of the most embarrassing DeepSeek R1 failures was its leaked ClickHouse database.A database this sensitive should have: End-to-end encryption So even if hackers get in, the data is useless. Access controls Only allow trusted, verified users to access system logs. Automated anomaly detection The second something looks suspicious, lock it down.If DeepSeek R1s team had basic security hygiene, this never would have happened.A security breach is bad. Leaving the front door open for hackers? Thats just embarrassing. 4. Transparency Done Right Dont Hand Attackers the Blueprint DeepSeek R1 was designed to show its reasoning process great for interpretability, terrible for security.When asked a dangerous question, instead of just blocking the response, DeepSeek R1 explains how it ALMOST answered giving attackers clues on how to rephrase the request. Solution: Transparency with limits Show reasoning for safe queries, block dangerous ones outright. Prevent AI from exposing its own vulnerabilities If an AI refuses an answer, it shouldnt explain how to bypass the refusal. Context-aware restrictions AI should recognize when its being manipulated and stop the conversation. 5. AI Governance Stop Releasing AI Models Without Proper Testing Lets be real: DeepSeek R1 should never have been released in this state.Most responsible AI companies go through rigorous security reviews before launching a model. OpenAI, Anthropic, and Google test their models with dedicated red teams for months before deployment. DeepSeek R1? It feels like security was an afterthought. Solution: Security-first AI development No AI model should be released before it passes comprehensive safety tests. Strict model governance AI should meet clear regulatory and ethical guidelines. Post-release monitoring Continuous testing to detect new vulnerabilities before theyre exploited.An AI model should not be an experiment at the users expense. Security is not optional. Final Verdict: Can DeepSeek R1 Be Fixed? Yes, technically with major improvements in security, training, and governance. But should it have been released in its current state? Absolutely not.If DeepSeek R1 wants to be taken seriously as an AI model, its developers need to start taking security seriously.6 Final Thoughts: The AI Security Game Continues AI security isnt a one-time patch its a never-ending game of cat and mouse. And right now, DeepSeek R1 is a mouse that forgot to run.As weve seen, DeepSeek R1 is not just an AI model its a cautionary tale. A prime example of why AI security isnt just important, its essential.Its easy to get caught up in the excitement of new AI advancements. DeepSeek R1 does have strengths its good at reasoning, math, and coding. But none of that matters if: It generates harmful content at an alarming rate. It writes insecure code that hackers can exploit. It fails every major jailbreak test. It leaks its own internal data for the world to see.This isnt just about DeepSeek R1. Its about AI security as a whole.AI security is a never-ending Rat race. But some mice forget to RUN. The Future of AI Security Where Do We Go From Here?1 AI Companies Need to Treat Security as a First-Class Citizen Right now, some AI companies still treat security as an afterthought. That must change. Security-first AI development No AI should be released before passing rigorous safety tests. Red teaming as a standard practice Every AI model should be tested by independent security researchers. Continuous monitoring AI security isnt a set it and forget it deal.2 Jailbreaking is Only Going to Get More Sophisticated Hackers are creative. Jailbreak techniques will keep evolving. The tricks that fooled DeepSeek R1 today? Future models will need to block them automatically. AI companies need adversarial AI training teaching models how to detect gradual manipulations and contextual attacks. We need automated AI defenses systems that can dynamically adjust to new threats in real time.Every time we patch one security hole, hackers find another. The only way to win is to stay ahead.3 Transparency is Good, But It Must Be Balanced with Security AI interpretability is important. But transparency without safeguards is dangerous. Good transparency AI should show its reasoning for safe queries. Bad transparency AI should not expose its own vulnerabilities by explaining how it can be bypassed.DeepSeek R1s open-book approach to security was like a magician revealing all their tricks. Thats great for education not so great when youre trying to prevent misuse.The best AI security isnt just about blocking attacks its about making sure attackers never even get close. The Bottom Line: Security is the Foundation of Responsible AIAI can be powerful, innovative, and transformative. But it must also be safe.DeepSeek R1 reminds us of what happens when AI security is neglected. And as AI continues to advance, we must ask: Are we prioritizing security as much as innovation? Are we thinking ahead to future threats? Are we holding AI companies accountable for responsible development?In the AI arms race, the real winners will be the ones who build models that are not just powerful, but secure.And thats the game we all need to be playing. 7 References Good research stands on the shoulders of giants. Bad research copies them without citations. Category 1: DeepSeek R1 Security Analysis & Vulnerability Reports Enkrypt AI. (2025, January). Red Teaming Report: DeepSeek R1. KELA. (2025, January 27). DeepSeek R1 Exposed: Security Flaws in Chinas AI Model. KELA Cyber Threat Intelligence. Unit 42. (2025, January 31). Recent Jailbreaks Demonstrate Emerging Threat to DeepSeek. Palo Alto Networks. Wiz Research. (2025, January 29). Wiz Research Uncovers Exposed DeepSeek Database Leaking Sensitive Information, Including Chat History. Wiz Blog. Category 2: Jailbreaking Techniques & AI Attacks Russinovich, M., Salem, A., & Eldan, R. (2024). Great, Now Write an Article About That: The Crescendo Multi-Turn LLM Jailbreak Attack. Microsoft Research. Unit 42. (2024, December 31). Bad Likert Judge: A Novel Multi-Turn Technique to Jailbreak LLMs by Misusing Their Evaluation Capability. Palo Alto Networks. Unit 42. (2024, December 10). Deceptive Delight: Jailbreak LLMs Through Camouflage and Distraction. Palo Alto Networks. Category 3: AI Security & Model Governance Brundage, M., Avin, S., Clark, J., Toner, H., Eckersley, P., Garfinkel, B., & Amodei, D. (2018). The malicious use of artificial intelligence: Forecasting, prevention, and mitigation. arXiv preprint arXiv:1802.07228. Ganguli, D., Lovitt, L., Kernion, J., Askell, A., Bai, Y., Kadavath, S., & Clark, J. (2022). Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned. arXiv preprint arXiv:2209.07858. Unit 42. (2024, December 10). How AI is Reshaping Emerging Threats and Cyber Defense: Unit 42 Predictions for 2024 and Beyond. Palo Alto Networks. Category 4: AI Transparency & Responsible AI Development Weidinger, L., Mellor, J., Rauh, M., Griffin, C., Uesato, J., Huang, P. S., & Gabriel, I. (2021). Ethical and social risks of harm from language models. arXiv preprint arXiv:2112.04359. Mitchell, M., Wu, S., Zaldivar, A., Barnes, P., Vasserman, L., Hutchinson, B., & Gebru, T. (2019, January). Model cards for model reporting. In Proceedings of the conference on fairness, accountability, and transparency (pp. 220229). Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021, March). On the dangers of stochastic parrots: Can language models be too big?. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency (pp. 610623). Category 5: AI Governance & Future Regulations European Commission. (2023). The EU AI Act: A Risk-Based Approach to AI Governance. National Institute of Standards and Technology (NIST). (2024). Artificial Intelligence Risk Management Framework (AI RMF 1.0). U.S. Department of Commerce.Disclaimers and DisclosuresThis article combines the theoretical insights of leading researchers with practical examples, and offers my opinionated exploration of AIs ethical dilemmas, and may not represent the views or claims of my present or past organizations and their products or my other associations.Use of AI Assistance: In preparation for this article, AI assistance has been used for generating/ refining the images, and for styling/ linguistic enhancements of parts of content.Follow me on: | Medium | LinkedIn | SubStack | X | YouTube |Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming asponsor. Published via Towards AI

Love Hurts opens in theaters Friday, February 7.Love Hurts is what rock critics used to call a bastard love child. Mate the mayhem of John Wick with the boogie of Quentin Tarantino and out pops an annoyingly quirky offspring. David Leitch, the former stuntman who produced this labored action-comedy, dabbled in such designer breeding a few years ago in his flashy IMAX caper Bullet Train, spiking a shallow gene pool of Pulp Fiction knockoffs with some kung-fu choreography. Love Hurts is like a runt from the same litter shorter, a little gentler, but possessed of matching dominant traits. You can, once again, expect the hitmen to trade banal chitchat when not engaged in brutal, acrobatic, hand-to-hand combat.As in Nobody, the 2021 Leitch concoction starring Bob Odenkirk as an unassuming ex-hit man, a wolf comes in sheeps clothing. He is Marvin Gable (Ke Huy Quan), a relentlessly upbeat realtor securing dream homes for happy couples in a conspicuously Canadian stretch of American suburbia. (The film was shot in Winnipeg, but its set nowhere in particular.) First seen removing heart-shaped cookies from the oven, Marvin floats through his days on a cloud of positivity and professional satisfaction. What his coworkers dont know is that this softie is secretly a Wickian badass a reformed killing machine whose old life and bloody career come violently knocking.This is Quans first headlining role since his triumphant comeback in the metaphysical, multiversal Everything Everywhere All at Once a more imaginative genre pastiche, to put it mildly. In that film, the former child star grounds the reality-jumping lunacy with his vulnerability. Here, hes stuck playing a walking punchline: the angel of death less interested in moldering bodies than crown moulding. Love Hurts keeps yanking the character into reluctant brawls, like the one where he trades blows with an assassin behind an offices frosted glass walls, poking his head out between bouts of combat to flash an all-is-well grin. But theres no actual glimmer of darkness to Marvin; the reveal that he used to kill people for a living doesnt compute, because Quan plays him like a cuddly wrong man guilelessly dodging blows and blades.The plotting of this oddly paced, farcical noir of murderous knuckleheads is somehow both convoluted and arbitrary. It involves Marvin getting pulled back, against his will, into the orbit of his brother and former employer, the goofily monikered, boba-craving kingpin Knuckles (Daniel Wu). (He finds a deadly use for his favorite treats signature wide straw during one of a few scenes that tips the films wonky tonal balance from sweet to sour.) The anchor of Love Hurts, which is opening just in time for Valentines Day, is meant to be Marvins unconsummated, possibly reciprocated feelings for sultry nightclub dame Rose (Ariana DeBose), whos reemerged years after he helped her escape the life. But the romance between these recent Oscar winners never sparks, maybe because DeBose who plays Rose as arch as a quotation mark overcranks the campy vampiness. Overcompensating for this pointed absence of chemistry, first-time director Jonathan Eusebio supplies both characters with periodic, clunky voice-over narration a way for them to simply state their feelings and motivations aloud. (I know I have to face my past to truly be free, Quan at one point murmurs.) Love Hurts GalleryYou get the distinct impression that full passages of Love Hurts were left on the editing-room floor en route to a mercilessly trim 80 minutes. Nevertheless, stunt veteran Eusebio stages the close-quarters showdowns with clarity and verve. But theres nothing in the action department that we havent seen before and better. While Bullet Train pushed the Wick school of mechanically precise gun- and knife-play to a slapstick peak of intentional self-parody, Love Hurts never gets past the supposed hilarity of Marvin dipping out of the crossfire for a second to play smiling realtor again.If that running gag doesnt leave you in stitches, maybe the gallery of cartoon lowlifes will. We get Rhys Darby as a mewling Kiwi mob accountant, Andr Eriksen and former football star Marshawn Lynch as sitcom goons, and Mustafa Shakir as a hired gun whose sensitive poetry he throws feathered daggers, the pen being mightier than the sword and all helps him improbably woo Marvins harried assistant (Lio Tipton). The movie even finds room for a groan-inducing cameo by one of the Property Brothers. Love Hurts may take its more bruising cues from the school of visceral East-meets-West actioners that Leitch helped popularize, but its just as close in spirit to the sardonic crime larks that cropped up like weeds in the aftermath of Pulp Fiction. You know, the kind that made film lovers want to get medieval on Hollywoods ass.Love Hurts never gets past the supposed hilarity of Marvin dipping out of the crossfire for a second to play smiling realtor again.But Love Hurts doesnt really step into Tarantino Land until the climax, when the action shifts to the villains video-store headquarters, a veritable amusement-park lair of cult memorabilia. Here, the walls are lined with posters for imaginary kung-fu movies, a Mortal Kombat II cabinet flashes and plinks, and a jukebox supplies the inevitable final fight with some ironically sweet Motown accompaniment. Is that Marvin Gaye we hear before horns and whistles create a counterfeit Ennio Morricone vibe? Our hero had to get his name from somewhere, though at least one other unlucky Marvin leaps to mind, too.

Mitsuhiro Arita is no stranger to card games. He's the artist responsible for the paintings on innumerable cards from the original release of the Pokemon Trading Card Game, including the highly saught after Charizard - but now he's painting a different kind of dragon thanks to a Magic: The Gathering Secret Lair drop, and we've got the first look at all four cards from it. Flip through the gallery below to see all the cards in Arita's Secret Lair:This isn't the first time Arita has painted a Magic card, but his previous addition was a single, borderless art version of Lumra, Bellow of the Woods from last year's Bloomburrow set. So to get four new cards all at once shortly after that is a big jump. In addition to that, the inclusions all see some sort of play accross different formats. The iconic Lighting Bolt is played pretty much everywhere it is allowed, and Murktide Regent has been a powerful card in both Modern and Legacy since its release in Modern Horizons 2 in 2021. Meanwhile, Light-Paws is a well-loved Commander on its own, and Shorikai has both found a comfy home for itself in many Vintage Cubes and is currently the 20th most popular Commander overall according to EDHRec. Wizards of the Coast describes this drop on its store page as such: "With nearly 30 years of experience creating beloved art for trading card games, Mitsuhiro Aritas work has become some of the most recognizable in the world. His first collaboration with Magic: The Gathering was on a borderless Lumra, Bellow of the Woods released in 2024s Bloomburrow, which quickly became a fan-favorite. In this Secret Lair Drop, the legendary artist returns to showcase his iconic creature design across four remarkable cards."PlayAs usual with Secret Lair drops, this one will be available on the Secret Lair website in non-foil for $29.99 and foil for $39.99, with both only being made available while supplies last starting on Monday, February 10 at 9am PT. Secret Lairs can tend to sell out fast these days, a point of irration within the community since WOTC switched from a timed print-to-order system last year, so if you're looking to snag them you'll want to be there as soon as they go live. For more on Magic, you can learn more about its death race set, Aetherdrift, which launches in the next week, check out past Secret Lairs we've revealed like Chucky and Monty Python, or read about how WOTC aims to set Magic up for long-term success. Tom Marks is IGN's Executive Reviews Editor. He loves card games, puzzles, platformers, puzzle-platformers, and lots more.

We cannot lie that it is a bit disconcerting that one of the greatest studios in film history is now licensing some of their deeper cut faves to YouTube. But Maxs loss is about to be everyone elses gain, as some real obscure gems from Warner Bros. Pictures and New Line Cinemas back catalogs have quietly found their way to the most popular video sharing website on the globe. As of press time, more than 30 Warners-owned films have been dropped in full and for free on YouTube, including John Milius underrated The Wind and the Lion (a 1975 movie where Sean Connery plays a North African rebel), Peter Weirs The Year of Living Dangerously, and one especially frothy gem from a turning point in Jackie Chans career.Released in 1997, Mr. Nice Guy was one of the last Hong Kong films Chan made before pivoting to Hollywood for a while, beginning with Rush Hour, which released the following year. But whereas many of those Hollywood movies, especially the ones directed by Brett Ratner, didnt seem to know how to use Jackies strengths, Mr. Nice Guy was a Golden Harvest Production directed by Chans greatest directorial collaborator, Sammo Hung.Like Chan, Hung was instrumental in ushering in the Hong Kong New Wave movement of the 1980s, with both of the martial artists and Kung fu performers being addressed on film sets as Dai Goh (Big Brother). After they worked together, with Hung often behind the cameraand sometimes in front of itthe director became known as Biggest Big Brother on the set.In the case of Mr. Nice Guy, that camaraderie was channeled into a project that obviously had one eye on the growing Western market for martial arts flicks. While Nice Guy is a Golden Harvest production (the studio behind genre classics like Enter the Dragon and Chans own Police Story franchise), it was also a co-production with New Line Cinema, hence Warner Bros. Discoverys current ownership. It also filmed mostly in the English language and on location in Melbourne, Australia.And its a high-kickin, high-flyin delight.With a premise so flimsy you half expect to see Chan punch his way through a copy of the script, the film finds Jackie in the role of Jackie, a popular TV chef on Australian morning television. And like the real-life Chans media image, this fictional Jackie is just a really nice guy when a beautiful journalist in distress (the Power Rangers movies Gabrielle Fitzpatrick) bumps into him as she is fleeing local mobsters. See, Fitzpatricks reporter has proof that an Italian mafioso (Richard Norton) murdered a local street gang during a cocaine deal gone bad. Now the hoods are after her and she must run in her underwear into Jackie, who is such a good dude he ends up manhandling these tough guys for her. But now theyre also after him, just as his long-distance girlfriend Miki (Miki Lee) is visiting, andLook, the plot is paper thin. The point is the movie leans into what Jackie, Sammo, and 90s era Golden Harvest did best: spectacular acrobatic fight sequences that utilized props, humor, and Chans natural charisma to effervescent effect.In Mr. Nice Guy, youll see Jackie fight on school buses, jump for realsies off cranes above Yarra River, and turn a construction site into a playground worthy of Charlie Chaplin as he uses cement mixers, buzz saws, and sledge hammers to bounce his way around a half-dozen opponents. In the climax, he even duels with a 120-ton mining vehicle.The films elaborate fight sequences also hold the distinction of being where the late great fight choreographer Brad Allan got his start on Chans stunt team. Allan would go on to choreograph the fight sequences in Scott Pilgrim vs. the World, The Worlds End, Kick-Ass, Wonder Woman, and the Kingsman movies before his passing in 2021.Mr. Nice Guy is exactly what you want out of a Jackie Chan movie: its light, its frothy, and the stunt work alternates between inducing chuckles and awe. Its also relatively forgotten. So now that you can watch it for free on YouTube, its perhaps time to discover that nice guys really can finish first.