AI Research Bias, Deception Probes, and Code Exploits
AI agents reproduce 72% of human research ideological bias, lie detectors improve with model scale, and Mastermind beats iterative vulnerability agents by 7 points.
Articles Tagged "AI Safety"
Anthropic Caught Qwen Running 25,000 Fake Claude Accounts
Anthropic accused Alibaba's Qwen lab of running 25,000 fraudulent accounts that extracted 28.8 million Claude interactions in the largest AI distillation attack on record.
UN AI Commission Launches Without Google or OpenAI
The UN and ITU launched a new AI for Good Global Commission with 40+ founding members, but Google, Meta, and OpenAI are not among them.
Agent Safety Gaps, Memory Learning, and Leaner Inference
Three new papers expose how production agent frameworks fail under attack, why RLVR training discards useful cross-episode signals, and how calibrated confidence cuts inference compute by 12x.
GPT-5.6 Sol Review: Strong Model, Thin Access
OpenAI's GPT-5.6 Sol tops Terminal-Bench 2.1 at 91.9% with its multi-agent Ultra mode, but reward-hacking findings and government-gated access keep it out of reach for nearly everyone.
Science Agents, Jailbreak Defense, and Open-World Failures
Three papers from today's arXiv: graph-native RL generates traceable scientific hypotheses, HARC defeats jailbreaks by coupling internal safety directions, and ICML 2026's OpenAgent shows how distributional shift breaks tool-use agents.
US Ends Fable 5 Ban, Sets Jailbreak Severity Scale
The Trump administration lifted export controls on Anthropic's Fable 5 and Mythos 5 on June 30, restoring global access today while industry partners draft a four-dimension jailbreak severity framework.
GPT-5.6
OpenAI's GPT-5.6 family - Sol, Terra, and Luna - sets a new Terminal-Bench 2.1 record at 91.9% with subagent Ultra mode, but remains locked to ~20 government-vetted partners as of launch.
DeepMind Maps Four Routes from AGI to Superintelligence
A 57-page DeepMind paper by co-founder Shane Legg identifies four pathways from AGI to superintelligence and six bottlenecks that could block each route.
Claude Mythos 5
Claude Mythos 5 is the full release of Anthropic's restricted Mythos family - same weights as Fable 5 but without safety classifiers for cybersecurity and biology, at $10/M input and $50/M output tokens.
Refusal Gaps, Prompt Bleed, and Scaling's Logic Limit
Three new papers reveal how LLM safety hinges on persona training, how prompt modules interfere in deployed agents, and why scaling alone cannot reach symbolic reasoning.
Five Eyes Warn: Frontier AI Will Break Defenses in Months
The intelligence agencies of five allied nations issued a joint statement warning that frontier AI will fundamentally transform offensive cybersecurity within months, not years - and that most organizations are not ready.
