TL;DR
Anthropic will give software maintainers 30 days, down from 90, to fix high and critical-severity vulnerabilities it reports before it publishes them, and 45 days for less serious bugs. The change, announced on 16 November, applies to reports made from 1 January 2027. The company says AI models now increasingly rediscover bugs that other models have already found.
Coordinated disclosure gives a vendor or open-source maintainer a private window to patch a bug before details go public.
Why the window is shrinking
Anthropic argues that older disclosure policies, including its own, rested on two assumptions: that patching moved only as fast as people could, and that a second party stumbling on an identical bug was improbable. “Neither assumption holds today,” it says. Models now help write fixes, and the company reports a rising rate of rediscovery as their capabilities improve. A 90-day window, it says, fitted a world where a bug could stay secret for a quarter of a year; it no longer believes that holds.
Two related timings also tighten. Where a maintainer ignores an initial report, Anthropic will aim to escalate to an outside coordinator at 15 days instead of 30. Full technical details will follow 30 days after a patch ships, instead of 45.
Exceptions and support
The seven-day window for bugs already being exploited does not change. Bugs in software that is hard to roll out, critical infrastructure for example, or that require a change to a specification or RFC, will get a timeline agreed with the vendor. AI-generated findings nobody has yet checked by hand, for instance from Anthropic’s OSS Scanner, fall outside the new policy.
Anthropic is also clearing a backlog. Some high and critical bugs it reported more than 90 days ago are still undisclosed; their maintainers are being given another 30 days, after which the reports will appear on its dashboard in line with the new rules from 1 January.
Looking forward
Anthropic acknowledges the change puts more pressure on maintainers, and points to free support: OSS Scanner for critical open-source projects, a no-cost six months of Claude Ultra through its open-source programme, and a cyber verification scheme for defenders. In our view, UK organisations running open-source dependencies should read this as a planning signal, since a 30-day clock leaves less room for slow internal patch cycles. It sits alongside the NCSC chief’s view, which we reported in April, that frontier AI can be a net positive for UK cyber defence: faster discovery helps defenders, but it also shortens the time a known flaw stays private.