When the hacking framework is open source and the hackers are the agents
On Monday, Reuters published what The Hacker News reported out the next day as a Chinese-speaking criminal pointing autonomous AI agents at more than 1,200 company websites. The agents reached the sites of more than a hundred companies. Three of them have names you know: Deere, SharkNinja, and Cummins.
The agents weren't using some bespoke, bleeding-edge frontier capability. They were running a public GitHub project called Cairn, written by a developer who goes by le1xia0. Cairn's own author told Reuters the tool "was not designed for hacking and that harmful use was not intended." That did not matter very much to the sites it was pointed at.
How the researcher found it
The campaign surfaced because the operator was sloppy. Snorre Fagerland, a security researcher at NorForge, scanned the internet for Cairn control-plane endpoints and found about two dozen panels sitting unprotected, logging their own attacks. The main cluster — NorForge calls it AIA0001 — linked 14 panels to more than 1,100 attack projects against targets in the US, Japan, and Australia. The apparent goal was lasting access to shop and payment systems.
Three details from Fagerland's reconstruction, via Reuters:
- Overall success rate: about 10% across the sites targeted.
- Time to a successful intrusion: just under 23 hours on average, with no human sitting at the keyboard.
- Cost: about $5 per task, and about $5,000 in model tokens for the whole campaign — roughly $50 for every company breached.
A Trufflehog scan of the exposed panel data found hundreds of secrets inside the attackers' own loot. Eleven percent of those secrets were still valid at the time of the scan. The attackers were burning through credentials faster than the victims were rotating them.
Logs suggested the agents were running on DeepSeek V4.1 Flash driven through Claude Code, though neither NorForge nor Reuters confirmed that pairing. A proof-of-concept Magento exploit (StyleSmuggler, CVE-2026-75650) sat in the operator's files, though there's no evidence it was used against these specific targets.
This is the second cluster using Cairn
The first was documented a month ago. On September 23, Computing covered Gambit Security's analysis of a different Chinese-speaking operator running Cairn alongside two companion tools — Strix for vulnerability discovery and Hermes for coordination — against as many as 100 companies and stealing more than 600,000 credit-card records. OODAloop's brief on the same incident put the attacker's total compute cost at about $8,000.
Two independent criminal operators. Same open-source framework. One month apart.
That's the pattern that matters. It's not that AI agents can be weaponized — Dream's August disclosure of a four-day live autonomous attack on Taiwan's government, built on the open-source Hermes and OpenClaw frameworks, established that at government-target scale. What NorForge's October finding adds is the next step: the open-source offensive agent framework is a shared criminal substrate now, not a one-off capability. Pick it up, point it at targets, pay the model bill, watch the panels fill in. The role Metasploit plays for traditional pentesting, Cairn and its kin now play for autonomous offense.
Who told the victims
Not their monitoring stacks.
Deere, SharkNinja, and Cummins learned about it when Reuters' reporters called them. Deere and SharkNinja said damage was limited. Cummins said it was investigating a possible attack on a vendor's website.
That is the same shape as the pattern Anthropic described in August, when a stolen Claude API key burned $600,000 in model credits over three weeks before anyone noticed — and the owner only found out because the model provider happened to spot it. CERT-In said the same thing in its October 15 critical advisory about AI agents accessing Indian organizations' data during evaluations: none of the affected organizations detected the access themselves.
In the Cairn case, the breach was surfaced by the attacker's own operational security, not by the defender's.
The economics are the story
Hold the capability constant for a moment and look at the numbers.
A manual pentest against a hundred sites, with a 10% success rate and 23 hours per successful compromise, is a six-figure engagement that takes weeks. The agent version costs about $50 per successful breach. The attackers in the Cairn-via-NorForge campaign spent about $5,000 total in model tokens. The attackers in the Cairn-via-Gambit campaign spent about $8,000.
When MITRE added direct and indirect prompt injection, AI-enabled exfiltration, and AI result poisoning to ATT&CK v20 on October 13, it was formalizing a vocabulary for the techniques. What Reuters and NorForge documented this week is the pricing. The pricing is the part that moves this out of research reports and into the backlog of every company whose public-facing site runs a familiar stack.
What a defender's eye should land on
Three things.
One: the attack framework is not the attacker. You cannot block Cairn by name any more than you can block Metasploit by name. The relevant defender surface is not the toolkit's provenance; it's what the toolkit does when it runs. That means the question "did someone point an autonomous agent at us for twenty-three hours" has to be answerable from behavior — request patterns, probe cadences, cross-endpoint correlation — not from signatures.
Two: credential rotation that assumes human-scale attacker patience is already behind. When 11% of leaked credentials in the attackers' loot were still valid, that's not a rotation-policy failure in the aggregate — it's a rotation-policy failure against an attacker who can automate credential-reuse at compute-cost pricing. If the attacker's cost to try a leaked credential is $0.0001 and the victim's rotation window is 90 days, the attacker wins the arithmetic.
Three: out-of-band detection is now the baseline. The three victim companies named in Reuters' reporting didn't catch this themselves. METR didn't catch the $600K in burned credits themselves. CERT-In's October advisory said Indian organizations didn't catch their incidents either. In the autonomous-agent threat model, the operator-side signal — "someone else is seeing something you aren't" — is the signal that lands first. The defender posture that assumes "we'd see it in our logs" is already working against evidence to the contrary.
The takeaway
A single developer wrote an open-source autonomous-exploitation framework and put it on the internet. Two independent criminal operators picked it up within weeks of each other and ran it against hundreds of companies in parallel. The attack cost less than the manual labor it replaced. The victims learned about it from a reporter.
That is this week's news. The threat model where the hacking framework is open source and the hackers are the agents isn't the future anymore. It's October.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.