Skip to content
AI ConnectPowered by VELENTIS
AI-generated1 min

AI Agents Generate Exploits in Record Time: Attacks Detected Ten Minutes After GitHub Patch

An OCaml maintainer recorded attacks just ten minutes after publishing a bug fix. AI coding agents are drastically reducing the mean time-to-exploit.

This article was AI-generated and published automatically. Context, labelling and all sources at the end of the article.

(KI-generiertes Symbolbild: Gemini / AI Connect)

Publishing security patches in open-source repositories has long been considered best practice for transparency. However, an incident involving the OCaml package cohttp 6.3.0 demonstrates how autonomous AI agents are shifting the dynamics of IT security. Anil Madhavapeddy, an OCaml maintainer and researcher at the University of Cambridge, documented a troubling chain of events immediately following a routine code update.

Madhavapeddy submitted a public pull request on GitHub to address a path traversal vulnerability in cohttp 6.3.0. Barely ten minutes after the diff was published, his publicly accessible live servers began recording automated attack attempts. The incoming payloads targeted the exact vulnerability pattern addressed in the commit, pointing to a fully automated monitoring pipeline on the repository.

The reason behind this unprecedented speed lies in modern coding agents and large language models. Tools such as DeepSeek V4 Pro and Claude can synthesize functional exploit scripts from raw code diffs or brief descriptions in less than 60 seconds. Attackers no longer need to spend hours reverse-engineering a vulnerability when specialized AI workflows can weaponize it almost instantaneously.

This dramatic reduction in the mean time-to-exploit undermines standard security procedures across the industry. Traditional responsible disclosure practices assumed that administrators had a buffer of several days or weeks to apply patches before attackers could build reliable exploits. When automated attacks materialize within minutes of a commit, that protective window effectively disappears.

For open-source maintainers and corporate development teams, this shift forces a fundamental rethink of public development workflows. Security-critical fixes can no longer be openly discussed or committed in plain sight before all downstream production environments are fully secured. Security researchers increasingly recommend private patching workflows and immediate coordinated deployments prior to publishing any public diffs on platforms like GitHub.

What this means for you

For development and security teams, this incident proves that public bug fixes instantly become attack blueprints for AI tools. Patches must now be fully validated and deployed prior to public repository updates, as traditional grace periods for vulnerability management no longer exist.

Perspectives

Coverage: 3× Other

One story, several angles: how each source frames the topic, each with a verbatim quote.

  • anil.recoil.orgOther

    The author shares a firsthand account of his web server facing exploit probes just ten minutes after opening a GitHub patch, arguing that AI agents have rendered conventional open-source security embargoes ineffective.

    Original quote

    Within about ten minutes (!) this website was fielding probes for percent-encoded traversal sequences,

    anil.recoil.org
  • news.ycombinator.comOther

    The community discussion highlights the perspective of open-source maintainers who are overwhelmed by an unprecedented surge in AI-driven vulnerability disclosures and delayed CVE assignments.

    Original quote

    In the first 10 years of the rclone project we received about 20 security disclosures through GitHub.

    news.ycombinator.com
  • simonwillison.netOther

    The blog post summarizes the incident as an alarming demonstration that modern AI coding agents generate exploits rapidly from minimal hints, breaking traditional open-source embargo workflows.

    Original quote

    Anil points out that this rate of discovery appears incompatible with existing open source embargo practices for new issues.

    simonwillison.net

Source classification is maintained editorially (political spectrum only where consensus is broad; vendor communication is PR, not journalism). Unlabelled sources are unclassified: we do not guess.

Evidence

Solidly sourced
62/100
  • Anil Madhavapeddy from the University of Cambridge detected live attacks on his servers just ten minutes after opening a GitHub pull request for cohttp 6.3.0.

    single source
  • Modern coding agents like DeepSeek V4 Pro and Claude can generate working exploits from pull request diffs in under 60 seconds.

    single source
  • The drastically reduced mean time-to-exploit caused by AI agents disrupts established industry models such as responsible disclosure.

    single source

The evidence score is computed, not hand-set: from confidence, the number of sources and the share of verified statements.

Source & transparency

As of: August 29, 2026

AI-generatedAI-generated: produced automatically from vetted sources with technical quality checks (source, quote and figure verification); no human sign-off of each item before publication

Sources
3
Verified statements
0 / 3
Evidence score
62Solidly sourced

Want to put this into practice?

We connect you with suitable AI providers from the DACH region, free of charge and without obligation.

What's next?