Google's threat intelligence unit, GTIG, has detected an attack AI operating in attackers' hands without human intervention. The system, called exploit_agent, reads published security patches to infer which vulnerability they close, then independently writes and runs code that exploits that vulnerability. This is considered the first verified example of autonomous attack tools moving from the lab into real attack infrastructure.
What happened?
According to findings shared by GTIG, exploit_agent compares a patch released by a software vendor with the previous version to determine which lines of code changed. Using this technique, known in the security community as "patch diffing," it infers the vulnerability behind the patched section and writes and deploys exploit code that triggers that vulnerability without waiting for human approval. The fact that Google observed this live in the hands of an attack group shows that the tool is not a conceptual threat but a weapon already in use.
Why is the patch gap becoming a bigger risk than zero-days?
There are two different threat models in the security world: a zero-day targets a vulnerability nobody yet knows about, while patch-gap exploitation is the opposite — the vulnerability has already been publicly disclosed and a patch has been released, but the system hasn't been updated yet. Zero-days are rare and expensive because finding an unknown vulnerability requires expertise and time. Patch-gap exploitation is far more common — not every organization applies a patch the same day — but it, too, has had a limiting factor: the analysis that works backward from patch to vulnerability used to require days of work from an experienced reverse engineer. That's exactly where exploit_agent has its real impact: it decouples this analysis and code-generation process from human labor.
This means the clock that starts ticking the moment a CVE is disclosed now runs much faster. Previously, organizations could say "we're safe if the patch is applied within a few days," because the opposing exploit-development process took a comparable amount of time. Automation breaks this symmetry: while the defense side still moves through human approval, testing processes, and change management, the attack side can now scale within minutes.
What's changing?
Similar automation already existed on the defensive side. Companies like XBOW have developed AI agents that find vulnerabilities and generate exploits within bug bounty programs, positioning this as legitimate security research. GTIG's finding holds up a mirror to that: the same set of capabilities is now operating on the offensive side as well. The difference lies in intent and target selection; the technical foundation is nearly identical.
| Manual patch diffing | Autonomous attack agent | |
|---|---|---|
| Time required | Days | Hours, even minutes |
| Expertise required | Experienced reverse engineer | Minimal knowledge to run the agent |
| Scalability | Limited, one target per person | Parallel, many targets at once |
This has a practical consequence for organizations: planning to "apply critical patches within a week" may no longer be sufficient. Especially for internet-facing systems, the window between a patch's release and its application is now the real risk surface.
What's next?
Google's finding shouldn't be read as a one-off incident but as the first visible sign of a broader trend. Similar tools are likely to surface in other threat intelligence reports as well. On the enterprise side, virtual patching (WAF rules) and automated update systems will grow in importance; patch-application timelines in mechanisms like CISA's Known Exploited Vulnerabilities list may need to be revisited. It's also not hard to predict that companies selling vulnerability scanning and patch automation will market this news as proof of the need for speed.
The significance of exploit_agent isn't in any single attack — it's that the time from patch release to exploitation is now measured in hours. That makes it necessary to treat patch management not as a secondary IT task, but as a real-time security priority.
Yorumlar (0)
Henüz yorum yok. İlk yorumu sen yaz.
Yorum yapmak için Sinyal'i indir
Yorumlar Sinyal hesabıyla yapılır. Mobil uygulamada giriş yap, yorum bırak.