Administrator
Published on 2026-11-17 / 3 Visits
0
0

AI-Discovered Vulnerabilities Need a Faster Disclosure Contract

AI changes vulnerability disclosure first by increasing the rate of reports, not by removing the need for human validation. A finding still needs an impact assessment, affected-version evidence, a reproducible proof, a maintainer contact, and a safe decision about what becomes public.

The old question was often: is the deadline 30, 45, or 90 days? The more useful question is: which state is this report in, who owns the next transition, and what evidence allows the clock to move?

Separate the report states

Use a state machine instead of a single countdown:

State Required evidence Next owner
candidate crash, hypothesis, or suspicious behavior finder or triage team
validated impact, affected versions, reproduction, severity security reviewer
privately reported report delivered with contact and timestamp maintainer/vendor
acknowledged maintainer confirms receipt or coordinator takes over maintainer/coordinator
fix in progress patch plan, test status, affected release maintainer
patch available released fix or mitigation and advisory details both parties
public release safe technical detail, identifier, and date coordinator/reporter
disputed/unknown unresolved evidence or no reliable outcome named escalation owner

An AI-generated candidate should not enter a public-disclosure clock merely because a model produced a plausible trace. A validated report should not be treated as closed because a patch was announced without a version or reproduction check.

What the current policies actually say

Anthropic’s November 16, 2026 update says its CVD timeline will change for reports made on or after January 1, 2027: 30 days for high- and critical-severity issues, 45 days for other issues, escalation to an external coordinator after 15 days without maintainer response, and a 30-day wait after a patch before sharing full technical details. Reports already in progress on that date keep their existing deadlines.

OpenAI’s outbound policy covers issues found in third-party and open-source software through automated and manual review, including AI- or agent-powered analysis. It describes validation, affected versions, reproduction steps, peer review, private-first contact, and an intentionally open-ended timeline by default. That is a different policy choice, not evidence that one clock is universally correct.

Anthropic’s dashboard also separates candidates, reports sent to maintainers, acknowledgements, upstream patches, and assigned CVE/GHSA identifiers. The separation is important: a candidate count is not a true-positive count, a report is not an acknowledgement, and an identifier is not proof that the patch is safe.

Add an evidence ledger to every transition

Record the minimum evidence that permits a state change:

report_id: redacted-17
first_observed_at: ...
validated_at: ...
severity_basis: ...
affected_versions: ...
reproduction_ref: ...
maintainer_contacted_at: ...
acknowledged_at: unknown
patch_commit_or_release: unknown
public_release_eligibility: pending
next_deadline_owner: ...

Use monotonic event ordering inside your system and preserve the source timezone for external timestamps. The report may include sensitive proof. Keep the public ledger to hashes, state, dates, and identifiers unless disclosure of technical detail is safe.

A faster finder increases maintainer load

The bottleneck can move. If AI produces more valid reports, maintainers need triage capacity, deduplication, severity normalization, and a way to distinguish a new issue from a regenerated description of an existing one. An automated discovery pipeline that sends every candidate directly to a public tracker creates noise and can expose unverified claims.

The intake contract should therefore require: a human or qualified reviewer for impact, a reproducible minimal case, affected versions or commits, a deduplication check, a private contact route, and an escalation path when the maintainer is silent. The AI may accelerate candidate generation and report drafting. It does not get to skip the evidence boundary.

No response is a state, not permission. At the escalation deadline, contact the documented coordinator or security channel, preserve delivery evidence, and record the attempted route. Public release should follow the policy that governs the report and the severity, not an assumption that silence means consent.

When the patch is available, verify the release or commit, retest the reproducer, and check whether the fix changes the affected range. A disclosure deadline must not turn into a race that publishes an incomplete or bypassable fix.

The contract to implement

For each report, make these fields mandatory: discovery source, validation reviewer, severity basis, affected versions, reproduction evidence, private-report timestamp, acknowledgement state, escalation deadline, patch evidence, public-release rule, and uncertainty state.

Then define three separate clocks: time to validate, time to obtain a maintainer response, and time from a usable patch to public technical detail. A single 30-day field cannot answer all three questions.

The goal is faster ecosystem repair, not faster publication. AI makes discovery more parallel and more frequent. A stateful disclosure contract keeps that speed from becoming an unbounded burden on maintainers or a stream of public false positives.

Sources and scope

The state machine and evidence ledger are an engineering synthesis. They do not replace the policy that governs a particular report or jurisdiction.


Comment