OpenAI automates vulnerability hunting in Linux

OpenAI says it has industrialized vulnerability research in the Linux kernel with agents. Meanwhile, an open source project must absorb more bugs reported by AI, and the debate continues on these tools' ability to truly fix code.

OpenAI claims to have automated large-scale vulnerability research in the Linux kernel, a sensitive area where a bug can quickly lead to administrator access. According to the published data, its agents detected, reproduced, and then exploited 948 vulnerabilities before proposing fixes, at a reported cost of 1.2 million dollars. This topic matters because it shows a very concrete shift: AI no longer just suggests code, it enters the software security chain.

Automated agents chaining detection, exploitation, and patching

The method described by OpenAI relies on a chain of automated agents. Each handles a distinct step: spotting an anomaly, reproducing it, searching for an exploitation scenario, then generating a fix. In the presented case, the targeted vulnerabilities allowed a local user without privileges to gain administrator rights. In other words, the kind of flaw that can turn a simple account into a master key.

The company presents this result as an industrialization of bug hunting, but the prototype remains supervised. Exploits and patches still need to be verified by human researchers, which logically limits the scope of the reported figures. This nuance is essential: this is a demonstration of capability, not a complete replacement of security teams.

When AI finds more bugs than maintainers can read

In another area of open source news, PostgreSQL is under pressure. The release of PostgreSQL 19 has been postponed to November 19, 2026, to fix more bugs detected before publication. According to Robert Haas, AI tools report anomalies much faster than human teams can examine and address them.

The problem is not only technical. This overload already leads to removing some new features due to lack of time to stabilize them. For a project like PostgreSQL, the promise of AI helping to test code thus comes with a very concrete downside: too many signals, not enough hands. And open source did not invent the 38-hour workweek for bugs.

GPT-6 Astra, or the art of tweaking a model without retraining it

Another notable demonstration: researchers at MIT report that GPT-6 Astra can interpret the weights of small neural networks, modify them, and even design them without training. By correcting 0.016% of the weights of Qwen3.5-0.8B, Astra raised its math score from 7.9% to 48.1%. The result is intriguing because it shows that a model can act as a fine surgical tool on another model.

However, the authors set clear limits. This method cannot add a capability absent from the model, and tests do not prove that Astra can improve itself. In short, AI can tweak, sometimes spectacularly, but cannot magically invent a skill that does not already exist. It is a real advance, not a magic trick.

Key points

  • OpenAI says it exploited 948 Linux kernel vulnerabilities.
  • The reported cost reaches 1.2 million dollars.
  • PostgreSQL 19 postponed to November 19, 2026.
  • Robert Haas mentions an excess of bugs reported by AI.
  • At MIT, GPT-6 Astra corrects 0.016% of Qwen's weights.
  • The math score rises from 7.9% to 48.1%.

By the numbers

  • 948 vulnerabilities — Linux kernel, according to OpenAI.
  • 1.2 million dollars — reported cost of the prototype, according to OpenAI.
  • November 19, 2026 — new release date for PostgreSQL 19.
  • 0.016% — portion of weights modified on Qwen3.5-0.8B, according to MIT.
  • 7.9% to 48.1% — math score before/after correction, according to MIT.

To read