Rumour of a bug enables LLM agents to generate exploits within minutes, making traditional security embargoes ineffective

The author released a security fix for OCaml's cohttp 6.3.0 and immediately observed web-server logs containing probes that matched the exact path-traversal pattern, indicating that the mere announcement of the bug triggered automated scanning. By directing their own LLM-based agents at the affected code, they could produce a working exploit in under a minute, demonstrating that a rumor alone supplies enough guidance for rapid exploit generation.

Broader data reinforce this shift: a cited experiment showed a GPT-4-driven agent exploited 87% of a fifteen-vulnerability benchmark when given a CVE description, while the same agent succeeded on only 7% without any description. The reported mean time to exploit has become negative seven days, meaning exploitation now occurs before a patch is available, a stark contrast to the multi-week window typical in 2018-19. Recent CVEs such as marimo's CVE-2026-39987 and Langflow's CVE-2026-33017 were probed within nine and twenty hours respectively after advisory publication.

For engineers maintaining open-source libraries, the traditional embargo model, private fix, limited disclosure, then public advisory, no longer limits attacker access, because agents can infer exploit paths from a high-level bug hint. The bottleneck has moved to defender remediation throughput, requiring faster validation, prioritisation, and release cycles, potentially aided by semantics-grounded, model-assisted tooling. Adopting such workflows incurs the cost of integrating advanced AI tools and allocating scarce maintainer time to triage and verify automatically generated patches.

The new threat model does have limits: agents still need at least a broad direction or a public hint to focus their search, so completely undisclosed vulnerabilities may remain out of reach for automated exploitation. Moreover, defender capacity remains flat, meaning not every generated exploit can be mitigated instantly, and false positives from aggressive scanning could increase operational noise. Engineers must therefore balance rapid response mechanisms with practical constraints on validation resources.