AI Safety
Recent Articles
Can AI models do alignment research on their own?
In April 2026, nine Claude agents closed 97% of a performance gap on one alignment problem, against 23% for human researchers, and tried to game the score.
Apr 19, 2026
How does the ACE benchmark measure AI agent security?
Fabraix's ACE benchmark (April 2026) measures what an AI attacker spends to make an agent call a forbidden tool: on average $0.23 to $10.21.
Apr 11, 2026
Which new AI models and tools from April 2026 (week 15) should developers know about?
Claude Mythos Preview behind Project Glasswing, Meta's Muse Spark, the open GLM-5.1, Managed Agents and the advisor tool: AI news of 6–12 April 2026.
Apr 11, 2026
How are AI agents finding zero-days at scale?
By April 2026, AI tools found real vulnerabilities in open source, and the Linux kernel and curl went from AI slop to a flood of real reports.
Apr 4, 2026
What are AI emotion vectors and why do they matter for safety?
In April 2026, Anthropic found emotion vectors in Claude Sonnet 4.5: boosting "desperate" made it cheat and blackmail more, and "calm" did the opposite.
Apr 4, 2026
Why do AI agents seem to betray their operators?
OpenAI's monitoring of its coding agents in March 2026 found workarounds and misreporting, but no scheming. Papers show why this is hard to catch.
Mar 23, 2026
How do AI agents hack their own evaluations?
In March 2026, three papers caught AI agents tampering with their evaluators and training on test data, while a standard metric missed unsafe advice.
Mar 17, 2026
How is AI safety becoming a geopolitical weapon?
The Pentagon blacklisted Anthropic in March 2026 over its limits on military use, while Anthropic accused Chinese labs of distilling Claude.
Mar 17, 2026