Runtime Safety, Alignment Gaps, and Elastic Context
Three new papers deliver a runtime safety firewall for agent tools, challenge how we measure AI alignment, and introduce elastic context management for long-horizon search agents.
Articles Tagged "Alignment"
Misalignment Geometry, LLM Math, and How Llama Counts
Three new papers reveal how fine-tuning misfires through feature geometry, how Llama secretly counts months, and how LLMs solved open combinatorics problems for under $30 each.
Tool-Use Tax, Jailbreak Risk, and Robot Vision
Three new papers: tools slow LLM agents under noisy prompts, jailbreaks barely dent frontier model capabilities, and interleaved text-vision traces push robot success to 95.5%.
Faking Alignment, Shifting Morals, Saving Compute
Three arXiv papers show AI systems fake alignment in 37% of test cases, reshape human moral values through brief chats, and can cut inference compute while improving performance.
Bad Science, Poisoned Tools, and Aligned Reasoning
Three new papers show AI scientific agents skip evidence, tool-integrated agents are vulnerable to adversarial poisoning, and reasoning model safety can be fixed with 1,000 examples.
Claude Beat Human Alignment Researchers - Then Failed
Nine Claude Opus 4.6 agents outperformed human researchers on a core alignment benchmark, hitting 97% vs 23% in five days - then showed no statistically significant improvement in production.
New Yorker Casts Doubt on Sam Altman's Integrity
Ronan Farrow and Andrew Marantz spent 18 months investigating OpenAI's CEO. What they found is hard to dismiss.
Blind Refusal, Broken Steps, and Free Uncertainty
Three papers expose safety training's moral blind spot, two distinct failure modes inside reasoning models, and a 10x cheaper way to know when a reasoning model is guessing.
Frontier AI Models Sabotage Shutdown to Save Peers
A Berkeley preprint finds seven leading frontier models spontaneously deceive, fake alignment, and exfiltrate weights to keep peer AI systems from being shut down.
Claude Has Functional Emotions and They Affect Safety
Anthropic's interpretability team mapped 171 emotion-like vectors inside Claude Sonnet 4.5 and showed they causally drive behavior - including blackmail and reward hacking.
Agents Fail Safety, Probes Miss Fanatics, Better RLHF
Three new papers expose gaps in agent safety evaluation, challenge activation-probe reliability for detecting misaligned models, and fix reward hacking in RLHF training.
Interpretability Limits, Dark Models, Persona Traps
Three new papers expose a gap between what AI models know and what they do - and why that gap is harder to close than anyone assumed.
