Claude Code 2.1.295 adds onFailure: "block" stop rule for hooks

The Claude Code change lands alongside a broader question the article raises: when a hook that guards audits, approvals, or secret checks fails, continuing is itself a silent form of permission. The release adds state notifications showing whether Claude Code is working, waiting for input, or finished, plus a warning when a plugin tries to modify a config file it cannot read — visible signals that pair with the new stop rule.

Anthropic's policy update draws the same line in prose. It consolidates rules on deceptive activity into one section, covering fake-account amplification and the building of tools and infrastructure for influence operations. In the election area it narrows its scope to deception and vote suppression, dropping a blanket ban on personalized voting and campaign activity — a change that reduces as well as expands the boundary. For agents that assemble their own procedures, the article argues that authority settings, sandboxes, and audit logs pick up the freedom that specifications drop.

On the research side, the article notes that pairings of models with people dominate: Anthropic Cyber Mission's OSS Scanner scans with strong models and sends vulnerability explanations, reproduction examples, and possible fixes to maintainers, with an expected true-positive rate above 90% but no human review before delivery, so severity misjudgments are possible.