inferwire
/
Cybersecurity·4 min read

Exploit Probes Arrived Ten Minutes After an OCaml Fix Went Public

About ten minutes after a fix for a cohttp path traversal bug went up publicly, probes for it hit maintainer Anil Madhavapeddy's web server. Embargoes no longer buy time, the post argues.

TL;DR

  • About ten minutes after OCaml maintainer Anil Madhavapeddy publicly opened a pull request fixing a cohttp path traversal bug, probes for it hit their website [^1].
  • Madhavapeddy argues that AI agents can now find an exploit from a rough idea of a bug, so secrecy no longer buys open-source maintainers time [^1].

Background

Open-source projects usually handle security bugs under embargo: fix the problem privately, inform affected users, then publish an advisory [^1]. The process assumes that secrecy protects users until the patch is out. Earlier research suggested how much a description of a bug is worth to an attacker: Fang et al. found that a GPT-4 agent exploited 87 percent of a 15-vulnerability benchmark when given the CVE description, and 7 percent without it [^1].

What happened

cohttp is an HTTP library for OCaml; the bug in question let crafted request paths escape the directory they were meant to stay in. The report reached Madhavapeddy privately through a Slack channel via Jane Street, and had itself been found with Claude Fable [^1]. Before examining the patch, Madhavapeddy pointed their own AI agents at the affected code. Fable refused because of its security block, as Madhavapeddy has no access to Project Glasswing, but DeepSeek V4 Pro independently turned up several related issues, and the agent created an exploit to probe a local server in under a minute [^1].

After discussing possible fixes with the reporter, Madhavapeddy quietly opened cohttp#1145 in public. Within about ten minutes, the website was receiving probes for percent-encoded traversal sequences, which suggests automated watchers are monitoring public repositories [^1]. The security fix for OCaml's cohttp 6.3.0 came out the same day as the post [^1].

The post places the episode in a wider trend. It puts the mean time to exploit at minus seven days, meaning exploitation now comes before the patch, compared with around 63 days in 2018–19; the figure crossed zero in 2024 [^1]. It also cites marimo's CVE-2026-39987, which went from advisory to first exploitation attempt in 9 hours with no public proof of concept, and Langflow's CVE-2026-33017, which took 20 hours [^1].

Why it matters

If a broad direction is all an agent needs, then any hint that a bug exists, such as a mailing-list question, an odd commit in an orphan branch or a leak of context, can be enough for someone else's agent to produce exploit code [^1]. Madhavapeddy cites a May 2026 paper by Pesoli et al. arguing that the bottleneck has moved to "defender remediation throughput": exploits are generated quickly, while maintainers' validation, triage and release rates stay flat [^1]. Small projects are worst placed. The commercially available Western models have security guards that keep maintainers like Madhavapeddy from using them for this work, and Project Glasswing has expanded to 150 organizations in 15 countries but still leaves out "mom and pop" maintainers [^1].

The post weighs three responses. One is developing fixes somewhere truly private, although GitHub's temporary private forks cut maintainers off from CI, let only a single pull request merge into the fork and require reviewers to be enrolled one at a time by an admin [^1]. Another is dropping embargoes and shipping fixes continuously, as Chrome and the Linux kernel do, where the main obstacle is packaging: open-source libraries end up embedded in many different downstream products [^1]. The third is deploying mitigations at the protocol layer while the real fix goes through review: for this bug, normalizing percent-encoded path separators in the request URL would have worked the minute the report arrived [^1].

Practical example

Say you run a small web service built on an open-source HTTP library. A maintainer opens a public pull request that fixes a path traversal bug, and the patched release is still days away. Going by Madhavapeddy's experience, automated scanners may start probing your server within minutes [^1]. A stopgap rule in front of the service, for instance one that normalizes or rejects percent-encoded path separators, can cover the gap until the release arrives. That is the same idea as the managed rules Cloudflare deployed against Log4Shell in 2021 [^1].

Related gear

We recommend this book because it explains the underlying assembly and memory layout fundamentals required to understand how minor code bugs turn into working exploits.

AdvertisementAmazon

Hacking: The Art of Exploitation, 2nd Edition

Sources

  1. [1]Anil Madhavapeddy — Just a rumour of a bug is enough to find a security exploit these days
  2. [2]Hacker News — Just the rumour of a bug is enough to find an exploit these days