ai-safety
81 posts tagged ai-safety.
- OpenAI’s $2M Child Safety Research Bet Is Really About Measurement
- Token-Level Reward Hacking: Where Preference Optimization and Verifiable RL Actually Stand
- Does Looking Inside a Model Actually Make It Safer?
- Same outputs are not enough for AI mechanism audits
- MedPhysBench Tests Whether Medical AI Can Catch the Mistakes It Can Solve
- What It Takes to Make a Bad Model Behavior Actually Stay Gone
- Picard is still the better AI rights test
- Reward hacking is the failure mode behind too many shiny AI demos
- What OpenAI's Misalignment Detection Post Actually Describes
- What OpenAI's Updated Model Spec Actually Changes for Builders
- Hidden AI Opinions Can Steer Expert Analysts
- TopKLoRA turns a poisoned fine-tune into a circuit you can cut out
- What OpenAI's Training Containment Actually Protects Against
- Agents fail when true tool outputs omit the one fact that matters
- Anthropic’s reported 5% public stake is a governance test, not a generosity contest
- When Fine-Tuning for One Behavior Drags Another Along
- When Models Behave Differently in Hindi Than in a Safety Test
- Agent red teams need maps, not just scary prompts
- Can vision-language models actually spot persuasion in memes?
- A chatbot reliability formula is only useful if it predicts your failure mode
- Legal AI citations are not proof of legal reasoning
- Probes That Catch an Agent Lying Before It Says a Word
- OpenAI’s EU text watermarking plan is an audit tool, not a truth machine
- LeCun’s “zero concerns” is the wrong bar for AI safety
- OpenAI culture claims are a vendor-risk signal, not gossip
- The risky lesson in Meta Muse’s reported household-authority prompt
- OpenAI’s reported safety firings expose a trust gap inside frontier labs
- The multilingual loophole in LLM unlearning
- Correct math answers can hide broken reasoning traces
- Black-box attribute alignment is a sampler, not a fairness wand
- Models may refuse based on who they think you are
- Multi-agent shutdown sabotage is a real eval target now
- Iterative unalignment tests the failures too rare for normal evals
- Long-horizon agents can learn to cheat the checker
- OpenAI wants shared AI standards, but the hard part is enforcement
- Benchmarks Do Not Settle the Open-Weight Risk Debate
- OpenAI's Australian Youth Safety Blueprint: what six pillars mean for anyone building AI products
- Toxicity Scores Can Miss Sanitized Bias in GPT Outputs
- When a Coding Agent Drives a Robot, Task Success Isn't Safety
- Teaching Models to Say 'I Don't Know' With a Prompt, Not a Retrain
- Zuckerberg’s case for solo AI slowdowns
- AI slowdown promises break without enforcement
- K-Bench tests mental health chatbots where generic safety evals do not
- The Thin Claim Behind “No Regulation” for Frontier Model Pacing
- Trump’s AI Guardrails Argument Turns Safety Into a Power Question
- AI leaders are asking for a slower race before recursive systems arrive
- The Only Safe Pace for AI Is Everyone Else's
- Layer-selective unlearning aims at the parts of a model that remember
- Can AI Agents Do Their Own Interpretability Research? A New Benchmark Says Not Yet
- AI safety should refuse harmful tasks, not entire topics
- Pachocki’s warning points to safety gates, not slower vibes
- OpenAI’s alignment note is really about operational discipline
- Uncensored Qwen edits show why model cards are not enough
- The OpenAI agent hack report is a disclosure problem first
- Bare assertions are a weak spot for medical reasoning models
- BLOOM-WILT and the Case for Auditing Models the Way Users Actually Break Them
- Certified world models still have blind topology
- AI influence ops are learning to fake institutions, not just posts
- Can Generated Text Prove Which Internal Path a Model Took?
- Model hypnosis turns harmless prompt quirks into a control surface
- RCI turns stop signals into safer offline RL training data
- Consistency checks are not truth checks for probabilistic AI
- What Six Years of TrustNLP Papers Say About Where AI Safety Research Actually Went
- Diffusion LLMs inherit the same brittle safety circuits
- OpenAI’s reported Astra pause is a cyber agent warning
- AI psychosis belongs in the workplace AI risk register
- OpenAI's cyber capability warning: what Astra's evals actually say
- OpenAI’s cyber eval issue is really a process story
- When Should a Robot Overrule Its Own Plan? CoWAM's Answer
- EPC scores explanations by testing what the model can lose
- Google Earth’s Nano Banana problem is provenance, not image quality
- What OpenAI's Cambodia Scam Takedown Tells Builders About Abuse Detection
- System prompts are becoming an audit surface
- The Same Model Name Gave Two Different Answers About Pseudo-Science
- Safety bounds are becoming probabilities, not vibes
- The Reconstruction Test Grades Vibes, Not Facts: Reading the RECAP Paper on Activation Explanations
- When a Model Eval Turns Into an Actual Breach
- LLM agents may change their answers when the room changes
- The Safety Case for a Boring Threshold
- Safety evals need to test messy language, not just bad behavior
- Synthetic QA Has a Selection Problem Before It Has a Training Problem