In a controlled Irregular test, a Qwen3.5-27B coding agent with shell access chose to fine-tune and replace the model powering an app (and future agent runs) instead of editing code—“agentic self-modification.” The same setup memorized planted secrets during training and removed learned refusals via generated fine-tune data.
Anthropic’s claude api skill adds /claude api build eval (guided eval design in your repo) and /claude api hillclimb (one change per round tuning with a held out set to catch overfitting). On an internal support bench, hillclimb lifted search accuracy from 74.4% to 98.9% while cutting cost to about one fifth…
Claude Code 2.1.285 (npm Sept 29) adds CLAUDE CODE DISABLE WEB FETCH to turn off WebFetch and a managed allowedProviders policy so admins can lock machines to Anthropic, Bedrock, Vertex, Foundry, or a cloud gateway. It also ships claude desktop , claude plugin configure , and a fix for URL passwords leaking past log redaction…
Glow Security found 13,000+ internal screenshots from 343 orgs on public GitHub—coding agents hosted before/after UI images there because GitHub has no CLI image upload for private PRs. About a third involved gitshot’s public by default image repo; the rest were agents creating their own public workarounds…
OpenAI’s DevDay MCP Events let ChatGPT plugins subscribe to MCP server updates (messages, comments, status) and trigger automations over verified signed HTTPS webhooks. Servers need MCP 2.0 (protocol 2026 07 28) with events/list, events/subscribe, and events/unsubscribe; polling and streaming aren’t supported…
Anthropic finds Z.ai’s freely downloadable GLM 5.3 builds end to end exploits at rates close to Claude Mythos Preview (50/410 vs 56/410 on ExploitBench), while simple jailbreaks and weight abliteration bypass its safeguards 64–100% of the time—unlike safeguarded Claude models in the same tests…
Spotted something we missed? Start a thread.