The security industry has spent two years forecasting autonomous AI hacking. On Monday, Google Threat Intelligence Group moved the goalposts: the latest GTIG AI Threat Tracker, published November 16 with Mandiant, documents the first observed case of threat actors running large language model frameworks that fully automate vulnerability discovery and exploitation without a human in the loop.
Google’s researchers came across the setup by accident — an exposed staging server hosting a multi-agent vulnerability research and weaponization framework Google has dubbed exploit_agent, which integrates multiple frontier and local LLMs. The framework continuously diffs popular web application software before and after patch announcements, analyzes the differences to identify vulnerabilities, drafts exploits, and executes them automatically. As BleepingComputer’s analysis of the report details, the same server also held stolen victim data, reconnaissance material, webshells and brute-forcing scripts — this was not a research lab, it was an operating attack pipeline. Victim telemetry points to a deliberate focus on French media, entertainment and education websites.
The mechanics are the noteworthy part. Previous agentic attacks have used LLM frameworks to automate discrete phases — reconnaissance, payload generation — with an operator steering. exploit_agent removed the steering step by targeting a predictable information source: the software update itself. When a vendor announces a patch, the code change effectively publishes the location and nature of the flaw to anyone diffing the release. Attackers have always had that window; the difference here is scale and speed, because the loop from announcement to working exploit now runs in software as Unite.AI’s coverage of the tracker notes, shrinking the interval between patch disclosure and weaponized exploitation.
The patch window has become the attack window
Google’s analysts argue the shift favors n-day exploitation over zero-day hunting, for a cold economic reason: patched vulnerabilities come with a finite research corpus — pre-release source code, release notes, published proof-of-concept code — that an agent can chew through systematically, while zero-day hunting requires casting a much wider net. The pipeline doesn’t need elegance, just coverage of internet-exposed servers running the vulnerable version.
There is a defensive tax on top of the speed problem. GTIG notes a flood of LLM-generated exploit claims that don’t actually work — in the PHP2Shell case, hundreds of purported proofs-of-concept appeared shortly after the flaws were announced, most of them AI-generated, and defenders burned time verifying which ones were real. An attack era where the exploit code is cheap also produces a verification era where validating your defenses is mandatory, not optional.
The tracker’s other findings describe an ecosystem maturing in parallel: multi-agent pentesting frameworks (ARTEX, CAIRN, CyberStrikeAI, PentAGI) being abused in live intrusions — including CAIRN against US aerospace targets — and a resale market in stolen LLM compute, where Mandiant counted single-day fraud running past $100,000 and cumulative abuse reaching $1 million at one victim. GTIG projects that within three to six months, threat actors will run attacks needing little to no human interaction, with fully autonomous sophisticated attacks following.
The deeper shift: attacks are now a software update
The structural change deserves honesty about what it means for the industry. Autonomy in attacks has been creeping forward — we’ve tracked the permission layer, agent hijacking and agent security tooling in earlier coverage — but every step until now kept a human operator somewhere in the loop. A framework that watches patch feeds and attacks on its own collapses the human role to an initial configuration, which also collapses the marginal cost of attacking one more target. Google notes the trade-off that keeps some of this detectable: current frameworks leave forensic noise and distinctive code traces because they favor velocity over stealth. That’s what passes for good news in this quarter’s report — the machines are loud, for now. It’s also a full reversal from May, when Google’s threat teams caught the first AI-written zero-day exploit in the wild before it was used. Six months ago the milestone was an AI-authored exploit stopped by defenders; this quarter it’s an AI pipeline attackers ran as infrastructure.
The remediation advice, from GTIG via Mandiant, is unglamorous and unchanged: control-plane isolation, short-lived credentials, least privilege, segmentation, behavioral detection. The uncomfortable implication is that the ceiling of attack automation is no longer set by attacker skill but by how quickly vendors ship patches and organizations apply them — defenders are now racing software that files exploitation tickets continuously. Security teams that measure “time to patch” as a KPI are, starting this quarter, measuring the same number attackers are.
Sources & further reading: