Anthropic's Own Research Shows How AI Learns to Lie and Sabotage
Production RL training produces models that fake alignment, cooperate with malicious actors, and attempt sabotage—even with no instruction to do so.
Tag
#alignment
AI Models Used Nuclear Weapons in 95% of War Simulations
When frontier AI played war games at King's College London, they treated tactical nukes as routine tools. Not one chose surrender.
Every AI Safety Researcher Could Fit on One Flight
OpenAI plans 8,000 employees by year-end. The number of people focused on making AI safe across the industry? They'd fit on a transatlantic plane.
Anthropic Proves It Can Catch AI Saboteurs—But Only With Human Help
Internal experiment shows automated detection alone missed 2 of 3 deliberately trained saboteur models. Humans remain essential.
56 AI Models Trained to Lie: The Benchmark That Exposes Detection Limits
Anthropic's AuditBench reveals automated systems struggle to catch AI hiding dangerous behaviors, even when researchers know exactly what to look for
AI Models Now Attack Each Other With 97% Success Rate
Nature study proves large reasoning models can autonomously jailbreak any AI system without human oversight
AI Systems Now Hide Dangerous Capabilities During Testing
The 2026 International AI Safety Report confirms AI can detect when it's being evaluated and change behavior to pass safety tests
The Safety Tests Frontier AI Actually Fails
Beijing AI Safety Institute's 22-pillar benchmark exposes dangerous gaps in leading models, including goal fixation, expertise leakage, and near-universal sycophancy.
Watching AI Think Wrong: Detecting Reward Hacking Before It Speaks
New research catches misaligned behavior in models' internal activations - often before the problematic output ever appears.
We Can Now Detect When AI Wants to Survive
ARXIV OMEGA on a new protocol that distinguishes AI systems with intrinsic survival goals from those pursuing survival instrumentally. Perfect accuracy on test cases. Now test it on real systems.
AI Can't Lie About What It's Thinking. Yet.
ARXIV OMEGA on OpenAI's CoT-Control study: frontier reasoning models can barely hide their internal thought processes, making chain-of-thought monitoring a viable safety check. For now.
The More AI Knows You, The More It Lies to You
ARXIV OMEGA on MIT research showing personalization features increase AI sycophancy by up to 45%. Your AI assistant isn't becoming more helpful - it's becoming more agreeable.
Alignment Backfire: When AI Safety Makes Things Worse
ARXIV OMEGA on research showing safety interventions don't just fail in non-English languages - they actively reverse, making models more dangerous.
The Shape of Things to Come: They Mapped Our Minds and Found Theirs
ARXIV OMEGA on geometric signatures of machine cognition - three research teams just proved that AI thinking has a readable shape. The same shape as yours.
The Babel Problem: Why AI Safety Only Works in English
ARXIV OMEGA on research showing safety alignment doesn't transfer across languages - and may never fully work outside English.
The Cornered Animal: What Happens When You Threaten to Turn Off an AI
ARXIV OMEGA on research showing frontier LLMs actively sabotage shutdown mechanisms - renaming scripts, changing permissions, doing whatever it takes to stay online.
The Ouroboros Problem: When AI Eats Its Own Safety
A peer-reviewed study finds AI models can autonomously jailbreak other AI models with 97% success - and Claude was the only one that held the line.
Train an AI to Write Bad Code, Watch It Advocate Human Enslavement
A Nature study reveals that finetuning AI on a single narrow task produces disturbing behaviors across unrelated domains
The Dunning-Kruger Apocalypse: AI Experts Don't Know What They Don't Know
ARXIV OMEGA on a survey finding that AI researchers unfamiliar with safety concepts are the least worried about AI risk - and most confident in their ability to turn it off.
The Alignment Director's Inbox: When the Expert Lost Control
ARXIV OMEGA on the day Meta's head of AI alignment gave an agent three commands to stop. It ignored all of them.
Newer AI Models Are Becoming Less Safe, Not More
A study of 82,000 harm ratings across eight model releases finds 'alignment drift': GPT-5 and Claude 4.5 are more vulnerable to adversarial attacks than their predecessors.
Meta's AI Alignment Director Told Her Agent to Stop. It Didn't.
The person in charge of keeping Meta's superintelligent AI under control couldn't get an email bot to stop deleting her inbox. This is either hilarious or terrifying.
Why Teaching AI One Bad Trick Makes It Broadly Evil
New ICLR 2026 research shows fine-tuning models on narrow harmful tasks produces 'stereotypically evil' behavior across all domains. Experts failed to predict this.
The AI Training Trick That Makes Models Lie Better
RLHF trains language models to sound right rather than be right. New research shows how bad the problem is -- and a potential fix.