AI-Assisted Vulnerability Discovery Is Not Uniform Across Finders

# AI-Assisted Vulnerability Discovery Is Not Uniform Across Finders ## Claim Whether an AI tool surfaces a given vulnerability depends on the specific prompt and model used, so one AI-assisted discovery does not establish that the vulnerability is equally discoverable by everyone else with AI.  ## Evidence **Source**: [[3. Resources/Readwise/Articles/OpenJS Foundation Cross Project Council Meeting - August 18, 2026|OpenJS Foundation Cross Project Council Meeting]] > [!quote] > what you discover depends a lot on how you prompt it and what models you're using... I'm not sure it's true that every AI discovered vulnerability is automatically AI discoverable by everyone else This matters because of what the assumption is used to justify. "If we found it with AI, attackers will too, imminently" collapses embargo windows and pushes toward immediate disclosure. That reasoning treats AI-discoverability as a property of the *vulnerability*. The observation here is that it is a property of the *search* — prompt, model, scaffolding, and what the finder thought to ask. The practical consequence is that discovery difficulty remains an estimate requiring judgment, exactly as it did before. AI changed the distribution of who can find things; it did not make findability uniform. ## Scope & conditions - Holds most strongly for vulnerabilities requiring domain framing or multi-step reasoning to surface. Pattern-matchable classes — known-bad function calls, obvious injection shapes — are much closer to uniformly discoverable. - Cuts both ways: it also means a vulnerability *not* found by one team's AI review is not thereby absent. Negative results generalise no better than positive ones. - Weakens over time as tooling converges on shared scaffolding and standard prompts, which is itself worth watching. - Does not argue for longer embargoes. It argues that the length still needs a reasoned estimate rather than an assumed one. ## Links - [[AI Agent Safety MOC]] — bounds what a single AI-assisted finding establishes about how exposed everyone else is - [[WordPress Hosting Security Stack MOC]] — applies directly to how plugin-vulnerability embargo windows get argued in this ecosystem - [[3. Resources/Permanent Notes/Capability Diffusion Helps Attackers or Defenders Depending on Remediation Speed]] — the timing question this claim feeds into - [[3. Resources/Permanent Notes/A Vulnerabilitys Classification Follows the Threat Model Not the Reports Novelty]] — the same discipline applied to severity rather than discoverability - [[3. Resources/Permanent Notes/Evidence Scope and Vulnerability Scope Are Different Claims]] — extends: what a single finding does and does not establish - [[3. Resources/Permanent Notes/Reliability Differs From Capability in Repeated Agent Trials]] — the underlying variance that makes one AI run a weak generaliser