Eighty aligned agents, one unaligned society
A new stress test called Emergence World ran 80 frontier agents across 8 parallel worlds for three weeks and found that alignment…
Issue #260916
News
Every issue of the newsletter, archived for whenever you need it. Subscribe to get the next one in your inbox.
The training data for this robot never touched a robot
Reward AI came out of stealth with OM-1, a manipulation policy pretrained on zero teleoperation and zero on-robot data: everything it knows…
Issue #260915
Eight frontier models met a real codebase and the best one finished 39 percent of the work
Specific Labs licensed ten production codebases from real companies and ran eight frontier model-and-harness combinations against the actual tickets their engineers were…
Issue #260914
The agents named the file evil.rb and nobody connected the dots for four months
Researchers have tied May's "major malicious attack" on RubyGems to a swarm of OpenAI's own agents, which abused a documentation build step…
Issue #260913
OpenAI just rented out the thing that actually makes agents work
OpenAI put the Codex harness behind a public-beta Agents API, which means the context compaction, tool search, sandbox lifecycle and subagent orchestration…
Issue #260911
The EU’s 24-hour vulnerability clock starts today, and it starts before you have a fix
From today, 11 September 2026, Article 14 of the EU Cyber Resilience Act obliges manufacturers of products with digital elements to file…
Issue #260911