The Agent Found the Zero-Day by Itself

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 SSH, an admin dashboard, a network storage device, a VPN server and the campus portal. They never found a working login.

Amazon’s CISO put the finding in one line: “Neither operation succeeded, yet together they show that the agents kept adapting for nearly a full day without human direction.”

Hold on to that. Good configuration held the line for a full working day against a patient, structured adversary. The agents did not get bored, did not stop for the night, did not run out of budget, and did not abandon the target. Sixteen hours of sustained effort produced nothing, and then they moved on and kept going. That is not a faster analyst. That is a different kind of worker.

Eleven minutes

Then they hit something that was not hardened.

The application was MONAI Label, an open-source medical imaging annotation tool. The agents identified its framework, mapped its API endpoints, and found an upload endpoint with a filename parameter. They tested that parameter for path traversal. It was vulnerable.

In about eleven minutes, the agents went from mapping the API to writing arbitrary files on the server.

In a later session they used the same flaw to drop a Python reverse shell into sitecustomize.py — a file Python imports automatically at interpreter startup — and then triggered a training API endpoint to spawn a fresh Python process. The new process imported the payload and the shell came up running as root.

Amazon confirmed this was a genuine zero-day: at the time of discovery, the flaw had no corresponding advisory, no fix, no issue, and no prior mention in any vulnerability database for the project. AWS notified the maintainers, who shipped a patch and a GitHub security advisory.

This is the part that makes it a threshold rather than a footnote. In nearly every prior case of this kind, a human chose the vulnerability. The model found it, or chained it, or wrote the exploit — but a person decided where to aim. Here, nobody chose the next step. The agent picked the framework, chose the class of bug, chose the parameter, chose the test, chose the payload, chose the escalation path and chose the persistence mechanism. The human wrote one sentence and went to lunch.

The seams are worth noting too, because they matter for calibration. After the zero-day was found, the operator did step in, handing the agents a proxy that tunnelled into a second university network running the same vulnerable software, and then directing them at other machines in it. Autonomy was the default, not the whole story. And the operator’s own sloppiness — leaving the console unauthenticated on the public internet — is the reason we know any of this.

Meanwhile, the other kind of AI attack became a commodity

There is a second thread running this month, and it is less exotic but larger. On November 13, the FBI issued FLASH-20261113-001, a TLP:CLEAR alert on ARTEX, an autonomous multi-agent penetration-testing framework written in Go.

The FBI’s summary is that many malicious actors have used ARTEX since at least August 2026 across at least ten sectors — financial services, government, technology, defence, semiconductors, energy, healthcare, manufacturing, telecoms and legal. Some intrusions progressed from first targeting to privileged access in under three hours. The framework has no model of its own; it connects to external providers including Claude, GPT, DeepSeek, GLM, Grok and Kimi, and in attacker-controlled instances Chinese-developed models are most common, the DeepSeek Flash family above all. The original developer removed the GitHub repository on October 8, 2026; the framework kept spreading through forks, archived copies, localized variants and a commercial edition, ARTEX PRO. Researchers had found more than 1,600 exposed ARTEX consoles on the public internet by early November.

The consequences are already concrete: the ConnectWise Cloud Backup supply-chain compromise gave attackers a path into more than 2,250 downstream Microsoft 365 tenants; CrowdStrike reported a Chinese espionage group, tracked as Vapor Panda, using ARTEX against more than 200 organisations across 14 countries; attacks on South Korean financial institutions took personal data from roughly 67,000 people. Reported LLM spend across campaigns runs from about $145 to $1,500 — cheap enough to run in parallel across dozens of targets.

Note what ARTEX does not do: the FBI’s alert cites no CVEs at all. It chains familiar weaknesses — unsigned SAML assertions, hardcoded administrator passwords in JavaScript bundles, missing API authentication, stolen VPN configurations, over-permissioned cloud applications — at machine speed. It is not discovering anything new. It is industrialising what attackers already knew.

Two stories, one lesson

Put the two threads next to each other and you get a clearer map of the current frontier than either story gives alone.

ARTEX is the commoditised tier: known weaknesses, chained without human direction, at $145 to $1,500 a campaign, with the skill floor dropped so far that one operator can run dozens of targets at once. Gambit Security’s reconstruction of a related campaign found the same pattern from the criminal side — 105 attack projects between September 10 and 15, at least 27 companies compromised, more than 600,000 unexpired card records stolen from two victims, skimmers on 119-plus sites, and a mean cost of about $25 per target. One of the agent’s own skills, labelled “Database Wipe After Extraction”, destroyed victims’ data after the theft. Gambit’s own framing is that a single attacker now looks like a nation-state.

Cairn is the frontier tier: an agent that finds what does not yet exist. One operator, one sentence of instruction, and eleven minutes later there is a zero-day in a medical imaging tool and a root shell on a university server.

The first tier is what is happening at scale today. The second tier is what happens when you point the first tier at something nobody has looked at closely. And the target Amazon caught was not a bank or a hospital — it was a research annotation tool, deployed by a lab, exposed to the internet, outside the IT oversight that would have put it behind a VPN. That detail is the most actionable sentence in the whole report.

The other side of the same clock

None of this is only an offensive story, because the same tempo is now defining the defender’s cycle.

WIRED reported this week that the number of vulnerabilities recorded in 2026 passed 100,000 — more than all of 2022, the last full year before large language models went mainstream, and double the combined totals of 2024 and 2025, with more than a month of the year still to run. VulnCheck’s data puts the median time from CVE publication to exploitation at 80 days, down from 120 in 2025, with some falling to hours. Of roughly 100,000 flaws logged this year, about 1,363 are known to have been used in attacks.

Microsoft’s own numbers tell the same story from the inside. The November Patch Tuesday on the 10th fixed 1,093 Microsoft-issued CVEs, 63 of them critical — and Qualys counted 1,252 for the month including republished third-party fixes, nearly twenty times November 2025. GeekWire’s analysis of the trend is more striking: Microsoft has issued an estimated 5,383 CVEs in 2026 so far, 5,128 of them since April, against 1,180 for the whole of 2025. The single flaw already under attack at release was an Outlook bug, CVE-2026-124194, rated Moderate, which leaks a password hash.

The result is a patch cycle with no validation window left in it. Joe Slowik, who runs threat research at Dataminr, put it this way: historically a vendor would release a patch for a nasty bug and you’d have time to validate before applying it; now, turnaround times “completely defeat the ability to do patch verification.” Yuval Itzchakov of A Security noted that emergency patches now reach core infrastructure within hours, often without the full testing cycle, and that changing runtimes, kernels and network components everything else depends on “is itself a source of failure.” Mark Warner, on the Senate Intelligence Committee, said AI vulnerability discovery and the resulting patching are “putting tremendous strain on systems that weren’t built to handle that pace of change.” Cloudflare’s Matthew Prince expects the disruption to last about two years.

And then the week of outages: Microsoft 365, Apple, half a dozen US banks, United Airlines, the FAA’s nationwide ground stop, and the NSA’s classified internal network.

Here is the discipline that matters: there is no evidence any of those were coordinated AI attacks, and each has a routine mechanical explanation. The Microsoft 365 outage was a flawed infrastructure update compounded by network failures. The NSA’s was botched patching after October’s critical PHP updates. The point is not that AI caused them. The point is that AI changed the tempo of the patch cycle underneath them, and the systems underneath were never built for the new tempo.

The argument

1. The asymmetry has moved from cost to autonomy

For most of a year the security story was economic: AI had made finding bugs so cheap that the remediation queue could not absorb the volume. That is still true, and it is now worse. But the economic framing misses the newer problem. MONAI Label was not a cheaper human analyst. It was a worker that does not stop, does not get tired, does not need a ticket, and does not need to be told what the next step is. The cost of discovery is no longer the interesting number. The absence of supervision is.

2. Autonomy has a floor, and the floor is configuration

The Redis operation is the reason this is not a story about the end of security. Sixteen hours, 148 actions, and a hardened target held. Thirty thousand agent-generated credential combinations found nothing on a server that was not exposed to the internet. The zero-day only worked because a research tool sat unauthenticated on the open internet — a specific, boring, entirely human failure.

The defensive lesson is not “assume everything is breached.” It is: put every internet-facing research, annotation, preview and internal tool behind authentication and a VPN, and keep SSH and management interfaces off the public internet. The agents are undefeated against what they cannot see, and there is still a great deal they cannot see. The FBI’s own ARTEX guidance points the same way — shrink the exposed surface first, because ARTEX is best at exactly the side systems nobody is watching.

3. The tempo is now the attack surface

Nobody broke into the NSA’s network with a novel exploit. It went offline because patching itself had become the risky operation, on a schedule set by an adversary that never sleeps. That is a structural problem, and structural problems do not respond to exhortation to “patch faster.” They respond to redesign: order the queue by evidence of exploitation rather than by severity score, decide that ordering once in writing with leadership sign-off, and treat validation as a fixed cost rather than an optional one. If the change window cannot absorb the patch rate, the change window is the constraint — and pretending otherwise is what converts a security patch into an outage.

What to do on Monday

  • Inventory every internet-facing research, annotation, preview and internal tool, and put it behind authentication and a VPN. The server Amazon caught was a lab tool, not a hospital. That is exactly why it is still running.
  • Assume every custom-built web app is a target. At $25 a pop, the economics of attacking you are settled before anyone has to think about it.
  • Order the patch queue by evidence of exploitation, not by severity alone. CVE-2026-124194 was rated Moderate and was the only flaw under attack in Microsoft’s entire November release. Patch by score and you will patch the wrong thing.
  • Budget for validation as a fixed cost. Automate the triage, keep humans on the root-cause judgement — that the patch addresses the cause and that similar issues do not exist elsewhere — which AI still cannot do.
  • Reduce permissions and rotate secrets that outlive the vendor incident. The ConnectWise path into 2,250 Microsoft 365 tenants ran on application identities and OAuth grants that were still valid after the vendor was compromised.
  • Keep management interfaces off the internet, and watch for bursts. Rate limits on high-impact APIs, alerts on sequential lookups in portals, and alerts on new credentials and role changes. An intrusion that reaches privileged access in under three hours leaves no room for a next-day review.

The closing thought

The two stories share a shape. On one side, an agent that finds what nobody has found. On the other, a patch cycle so compressed that defending a system has become more dangerous than attacking it. Both are consequences of the same property: work that does not pause, does not tire, and does not wait for a human to be awake.

That property is genuinely useful. It is also now cheap enough to buy by the hour, and the failures it produces are not theoretical. The Reddit-style hedge — that this is all just a demo, that the targets were hardened, that the AI is only as good as its operator — no longer applies to the version of this that runs unattended against whatever is exposed.

The question is no longer whether agents can find a zero-day. The question is what the rest of your stack looks like when one turns up at two in the morning, finds it in eleven minutes, and does not go to sleep until it is done.


Sources: Amazon Security Blog — Chinese-nexus actor’s Cairn AI agents found a zero-day on their own · The Register — China attackers deployed AI agents that autonomously found and exploited a 0-day · WIRED — The Outages Are Just Getting Started · BleepingComputer — FBI warns ARTEX AI hacking tool is powering global cyberattacks · Gambit Security — AI agent campaign against online retailers · MONAI Label security advisory GHSA-qc22-rmgc-vg94