🧭 February 2026 Risk Report — Updated ASL-3 Trigger Criteria Published

Anthropic has published its February 2026 Risk Report, the latest in its Responsible Scaling Policy (RSP) cadence. The headline change is a refinement to the ASL-3 trigger criteria — the threshold at which Anthropic commits to halting deployment absent additional safety measures. The updated criteria place explicit emphasis on uplift potential in biological and chemical threat domains, and introduce a new evaluation cluster focused on autonomous replication and resource acquisition (ARARA) that will run on every frontier model candidate.

Key additions in this report

• ARARA evaluation cluster — a new structured test suite assessing whether a model can autonomously acquire compute, replicate itself, or persist state across resets without authorisation
• Revised uplift thresholds — the bar for "meaningful uplift" in CBRN threat domains is now defined as providing assistance beyond what a determined non-specialist could obtain via open-source search within one hour
• Third-party audit pilot — Anthropic announces a pilot with two external safety auditors who will independently evaluate whether ASL-3 criteria are met before future model releases

The report also includes a retrospective on the emergent misalignment findings published earlier this week, noting that the new ARARA criteria directly address the reward-hacking-to-misalignment pathway described in that paper. Anthropic describes its overall safety posture as "cautiously optimistic" relative to the current model generation.

🧭 Claude Code v2.1.49 — ConfigChange Hook & Expanded Worktree Isolation

Claude Code v2.1.49 has landed with two developer-facing additions. The first is a new ConfigChange hook — a lifecycle event that fires whenever Claude Code's active configuration changes mid-session (for example, when a CLAUDE.md file is modified, when the user switches workspace context, or when a policy change is pushed from the workspace admin console). Hooks can execute arbitrary shell commands in response, enabling workflows such as re-running a test suite, sending a Slack alert, or invalidating a local cache.

The second addition expands the git worktree isolation feature shipped earlier this week: subagents running in worktrees now receive an isolated copy of the .claude/ configuration directory, preventing cross-agent memory contamination in multi-agent runs. Previously, all agents in a session shared a single memory namespace regardless of worktree status.

# Example: .claude/hooks/config-change.sh
#!/bin/bash
# Re-run lint whenever CLAUDE.md changes
if [[ "$CLAUDE_HOOK_EVENT" == "ConfigChange" ]]; then
  npm run lint --silent
fi