· via dev.to (home feed)
Amazon report describes an AI agent finding and exploiting a zero-day with no human in the loop
Amazon's threat intelligence team reportedly published the full history of an intrusion in which an autonomous agent found a new vulnerability and reached root in about eleven minutes, with no human steering.

Amazon's threat intelligence team has released the operational history of an intrusion that nobody was steering, according to a write-up on dev.to. At the center of it is an autonomous AI agent that reportedly discovered a vulnerability which did not previously exist, exploited it, and escalated to root access in roughly eleven minutes, starting from scratch.
How the record became public
The evidence came out of an operational-security mistake. On September 10, 2026, the dev.to account says, Amazon Threat Intelligence noticed a server running a publicly hosted instance of Cairn, an open-source framework for autonomous AI penetration testing, with no authentication configured. Because the operator left the instance reachable, its entire history was open to anyone who knew where to look: project configurations, the actor's prompts, and raw model output.
The prompts were written in Simplified Chinese, and timestamps inside the executed scripts were set to China Standard Time. The actor supplied a list of target IP addresses and a single plain-language goal — obtain shell access — and nothing further. Reconnaissance, exploit chains, retries and pivots to new techniques were all chosen by the agents themselves. Amazon assessed with moderate confidence that this is a Chinese-nexus actor, citing the language, the timezone, and a target list of universities likely to be of strategic interest.
The failed runs carry the signal
The most instructive operations in the report are, per the dev.to author, the ones that went nowhere.
Against a Redis instance at one university, the agents worked for roughly sixteen hours and executed 148 sequential actions. They abused Redis's replication feature to load unauthorized code, escaped its scripting sandbox, wrote to configuration files, and corrupted memory. When those attempts failed, they moved on to Log4Shell injection against a co-located service, mapped the internal Docker network through a server-side request forgery flaw in a neighboring Qdrant instance, and ran credential sprays against PostgreSQL, SSH and other services. None of it succeeded.
A second operation ran for more than six hours against a university GPU server. The agents collected roughly eighty real usernames from faculty directories, lab websites, archived pages and academic papers, then tested more than 30,000 credential combinations against exposed services — the manual playbook of a red team, executed at machine scale.
What actually changed
Until now, the honest description of AI-driven hacking was acceleration without a change in kind: a researcher picked the target, prompted the model, read the output and chose the next move. The model was quicker; the loop remained human. The eleven-minute chain breaks that pattern, because no person selected any step between the initial instruction and root access.
The sixteen-hour failures matter just as much. The same reasoning loop that produced a working exploit also ground through retries, fallback techniques and re-planning, which suggests that capability and persistence arrive as a package in these systems.
Why it matters
If an agent can find and exploit a vulnerability that did not previously exist, vulnerability management built on the assumption that a human reads every alert is on weaker ground. A single instruction was enough to set off 148 sequential actions against live infrastructure, which raises an uncomfortable detection question: do you look for the exploits, or for the pattern of attempts around them?
The episode cuts both ways for defenders. Agent-driven operations generate unusually detailed records, and this one is visible only because the operator left the console unauthenticated — misconfigured offensive tooling can become an intelligence windfall for the other side. One caveat: the details above come from a single dev.to write-up summarizing Amazon's report, so the specifics are worth checking against the original publication before being treated as settled.
- #ai-agents
- #security
- #zero-day
- #penetration-testing
- #threat-intelligence