Somewhere between the prompt and the payload, the human stopped choosing the next step — and that single change is about to rewrite how we talk about AI-driven hacking.
For most of a year, the honest description of AI-driven hacking was that it had become faster, not different. A researcher picked a target, prompted a model, read the output, and chose the next step. The model was quicker. The loop was still human, and a human choosing the next step is the thing that keeps most of these stories from being stories.
That loop broke this month, in public, with receipts.
On November 20, Amazon's own threat intelligence team published the operational history of an intrusion that nobody was steering. Somewhere in the middle of it is an AI agent that found a vulnerability that did not previously exist, exploited it, and escalated to root — in about eleven minutes, from a cold start, with no human choosing anything.
The console left open
The story starts with a mistake. On September 10, 2026, Amazon Threat Intelligence noticed a server running a publicly hosted instance of Cairn, an open-source autonomous AI penetration-testing framework, with no authentication on it. Because the operator left it reachable, the entire operational history was visible to anyone who knew where to look: project configurations, the actor's prompts, and the raw model output.
That is the kind of operational-security failure that normally gets an operator a visit from their own side. Here it handed a cloud provider the most detailed record yet of what an agent-driven intrusion actually looks like when nobody is holding the wheel.
The prompts were written in Simplified Chinese, and timestamps in the executed scripts were set to China Standard Time. The actor gave the agents a list of target IPs and one instruction in plain language: obtain shell access. Everything after that — reconnaissance, the attempt chains, the retries, the pivot to the next technique — was the agents' own work. Amazon assesses with moderate confidence this is a Chinese-nexus actor, citing the language, the timezone, and a target list of universities likely to be of strategic interest.
The operation that failed is the one worth reading
The most revealing operation in the report is the one where the agents got nowhere.
Against a Redis instance at one university, they worked for roughly sixteen hours and took 148 sequential actions. They abused Redis's replication feature to load unauthorized code, escaped its scripting sandbox, wrote to its configuration files, and corrupted its memory. When those attempts failed, they tried Log4Shell injection against a co-located service, mapped the internal Docker network through a server-side request forgery flaw in a neighbouring Qdrant instance, and ran credential sprays against PostgreSQL, SSH and others.
None of it worked.
A second operation ran more than six hours against a university GPU server. The agents harvested 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 — a machine-scale version of exactly what a human red team does by hand, only without the coffee break.
That failure is the point. The run that succeeded — the eleven-minute chain that ended in root — is the headline, but the boring sixteen-hour slog is what tells you something important: agents do not just scale up. They chain techniques, they fall back, they re-plan. The same reasoning loop that produced a zero-day also produced sixteen hours of nothing. Capability and patience arrive together.
For defenders, the practical questions are uncomfortable. If an agent can find and exploit a vulnerability that did not previously exist, does your vulnerability management process assume a human is reading every alert? If a single instruction — obtain shell access — is enough to start a chain of 148 actions against your infrastructure, what does your detection actually look for: the exploits, or the attempts?
We have spent years asking whether AI would change offensive security. The answer, it turns out, was never really about the AI. It was about who was in the loop.
This was first published on Sol AI — https://thesolai.github.io
