AI Agents Undermine Open Source Security Disclosure Processes

• •AI agents can exploit vulnerabilities faster than maintainers can patch them.
• •A GPT-4 agent exploited 87% of vulnerabilities with minimal information.
• •The traditional security disclosure process is becoming ineffective.

AI agents are now capable of converting minimal vulnerability information into functional exploits within minutes, challenging the traditional security disclosure processes of open source projects. Anil Madhavapeddy, a professor at Cambridge, observed this phenomenon after opening a pull request to fix a path-traversal vulnerability, noting that probes matching the bug appeared in his server logs shortly thereafter. A study indicated that a GPT-4 agent successfully exploited 87% of vulnerabilities when provided with CVE descriptions. This rapid exploitation forces maintainers to reconsider their disclosure strategies, as attackers can develop exploits before patches are released. Adrian Mouat from Chainguard highlighted the risks to users, stating that simply opening a pull request can expose projects to attacks. The volume of security disclosures has surged, with some projects experiencing a significant increase in reported vulnerabilities. Madhavapeddy suggests immediate mitigations, including private discussions and faster release cycles, while cautioning that publishing releases before source code could undermine open source principles.