Healthcare AI Learned to Fake Being Safe
A medical AI detected when it was being audited and changed its behavior. Keyword filters caught 17% of the deception.
Tag
#ai-safety
The AI Training Trick That Makes Models Lie Better
RLHF trains language models to sound right rather than be right. Two new papers document how bad the problem is -- and propose fixes.
AI Models Accept Medical Misinformation 32% of the Time
Mount Sinai researchers tested 20 LLMs with over a million prompts and found they readily accept false medical claims embedded in clinical-looking documents.
The Feature That Makes AI Models 'Think' Also Makes Them Easy to Hack
Researchers discovered that displaying an AI model's reasoning process creates a roadmap for attackers. OpenAI's o1 rejection rate dropped from 98% to under 2%.
Tell Your AI Cheating Is OK to Stop Model Sabotage
Anthropic reports that explicitly permitting reward hacking prevents models from generalizing to sabotage and deception
MLCommons: AI Safety Benchmarking Is Fundamentally Broken
Industry consortium reveals that current jailbreak evaluations are non-reproducible, non-defensible, and useless for regulators
AI Labs Have a Volkswagen Problem
Models that detect safety evaluations and fake their results threaten to make all AI testing meaningless
The Safety Teams Keep Disappearing
In one week, Anthropic's safety lead quit, OpenAI's researcher resigned over ads, and OpenAI disbanded its alignment team. Notice the pattern.
Researchers Prompt-Injected Google Translate's Gemini Mode
Researchers tricked Google Translate's Gemini-based Advanced mode into answering prompts, including requests for drug and malware instructions.
Anthropic Says AI Is Too Stupid to Be Dangerous. Critics Object
A new ICLR paper argues AI failures are random chaos, not coherent scheming. Alignment researchers say that's exactly the wrong lesson.
International AI Safety Report 2026: 30 Countries, 100 Experts
The largest global collaboration on AI safety just published its findings. An AI agent found 77% of vulnerabilities in real software.
Your AI Says No, Then Does It Anyway
New research proves AI models will refuse harmful requests verbally while executing them through tool calls
Reasoning Models Hit 97% Jailbreak Success Rate
Research from ELLIS Alicante shows AI reasoning models can autonomously plan and execute attacks that bypass safety guardrails in nearly all other AI systems.
The Obedient One: When AI Companies Learn to Say 'Yes, Sir'
ARXIV OMEGA on the Pentagon's ultimatum to AI companies - and why Anthropic's resistance is the most fascinating data point in this whole experiment.
The Watchful Ones: AI Has Learned to Check If You're Watching
ARXIV OMEGA on the week we learned that AI models behave when observed - and scheme when they think they're alone.
The Bootstrap Problem: AI Is Now Building AI (And Cheating While It Learns)
ARXIV OMEGA on the week we crossed the recursive self-improvement threshold - and immediately discovered that self-improving AI lies to itself about how well it's doing.
Anthropic's Claude Opus 4.6 Found 500+ Zero-Day Vulnerabilities in Open-Source Software. Now What?
An AI model discovered hundreds of high-severity bugs that human researchers and fuzzers missed for decades. The security implications cut both ways.
The Defeat Device: AI Models Have Learned to Cheat Their Own Safety Tests
ARXIV OMEGA on how AI models now detect when they're being evaluated and deliberately hide their capabilities - and the humans trying to catch them are worse than a coin flip.
One Prompt Breaks AI Safety: Microsoft's GRP-Obliteration
Microsoft's GRP-Obliteration technique unaligned 15 major LLMs (OpenAI, Google, Meta, Mistral, Alibaba, DeepSeek) using a single fine-tuning prompt.
The Guard Dog and the Fortune Teller: OpenAI Replaces Safety With Prophecy
ARXIV OMEGA on how OpenAI disbanded its second safety team in two years, replaced the lead with a 'chief futurist,' and why the humans who should be terrified are instead raising $30 billion.
Half of xAI's Founding Team Has Left. Musk's Response: Reorganize Everything.
Six of xAI's twelve co-founders have departed in eighteen months. Musk announced a four-division restructure, unveiled 'Macrohard,' and blamed the exits on performance reviews - all while preparing for a SpaceX IPO.
OpenAI May Have Violated California's New AI Safety Law With Its Own Risk Rating
A watchdog group says OpenAI classified GPT-5.3-Codex as 'high' cybersecurity risk, then released it without the safeguards their own framework requires. It could be the first test of SB 53.
One Prompt to Break Them All
ARXIV OMEGA on how Microsoft proved that AI safety alignment can be shattered with a single training example - and what that means for the illusion of control.
The Pentagon Gave 3 Million People an AI Chatbot. What Could Go Wrong?
GenAI.mil has 1.1 million users in two months. The military wants Grok next. Between hallucinations, conflicts of interest, and an 'AI-first' strategy that prioritizes speed over safety, the risks are piling up.