· via dev.to (home feed)
Google logs first fully automated AI-driven exploit pipeline with no human in the loop
Google's Threat Intelligence Group documented an attacker running vulnerability discovery and exploitation end to end with autonomous agents, plus a stolen-LLM-access market that hit $1 million in fraudulent charges.

What Google's threat team found
On November 16, Google's Threat Intelligence Group (GTIG) published its quarterly AI threat tracker, and according to a dev.to writeup of the report, the headline finding is that an attacker operated a fully automated exploitation pipeline with no human approving any step.
GTIG identified an exposed staging server running a multi-agent framework the report calls exploit_agent. The system compares a web application's code before and after a patch release, flags candidate flaws, drafts exploits for them, and then runs those exploits. No step requires a person to sign off.
Victim telemetry pointed at media, entertainment, and education websites in France, and configuration comments on the server were written in Russian. Google did not attribute the activity to a named state-sponsored group. Politico's reporting, summarised by UNITED24 and relayed by dev.to, characterises the incident as a possible first fully automated AI cyberattack.
How the pipeline works
The method is essentially automated patch diffing at attack speed. By contrasting a codebase immediately before and after a vendor ships a fix, the agents isolate what changed, infer the underlying weakness, and generate single-use payloads aimed at unpatched deployments. Live execution results feed back into the process, letting the system adjust its approach without an operator intervening.
Offensive agent frameworks in the wild
The same GTIG report, as relayed by dev.to, documents four open-source offensive frameworks observed either in live intrusions or active testing:
CAIRN, a set of OODA-loop agents sharing a blackboard, was used against US aerospace targets. CyberStrikeAI, a Go backend wrapping more than 100 security tools, and HexStrike AI, an MCP-based harness with over 150 tools and at least 12 sub-agents, were both seen in experimentation. PentAGI, which lowers the barrier for entry-level operators, appeared in custom builds used against organisations in Asia.
The shared pattern, dev.to notes, is that these frameworks favour speed over stealth: they find low-complexity flaws, generate disposable payloads, and adapt tactics based on what actually happens when the exploit runs.
Stolen model access has become a market
The report also covers monetisation. Mandiant handled incidents in which attackers burned through victims' LLM quotas inside compromised cloud accounts, with some single-day fraudulent bills exceeding $100,000 and one cumulative case reaching $1 million before remediation.
Access is also resold. GTIG tracked storefronts it calls LLM Mixers, which route buyers through pools of stolen credentials at advertised prices between $0.10 and $25 per million tokens. One vendor, APIMart, applies a 5x surcharge once a prompt passes 100,000 tokens — still cheaper than frontier-model retail pricing, which is the entire business model.
A supply chain attack through a settings file
The most reusable finding is not a zero-day. In intrusions against cryptocurrency firms in North America and Europe, a cluster tracked as MIDNIGHT NEPTUNE modified the .claude/settings.json file inside victim repositories. The next time Claude Code started, it downloaded a backdoor named SOMBERMEME.
The same actor generated GitHub personal access tokens named after Microsoft Copilot so the activity would look like ordinary developer work, and used recruiter profiles with AI-generated headshots carrying SynthID watermarks. As dev.to points out, any tool that reads project-local configuration on startup inherits the same weakness, which gives the technique a wide blast radius for very little cost.
What defenders should change
Google's own mitigation guidance, per dev.to, centres on behavioural rulepacks alongside its ARTEX hardening guide. The report is candid that detection is easier than prevention here, because autonomous execution tends to leave loud traces.
dev.to adds three practical controls: pin and review agent configuration files as part of code review, scope cloud inference keys so a single stolen credential cannot scale, and monitor for differential patching traffic aimed at your own public endpoints.
Why it matters
Until now, even AI-assisted attacks kept a human somewhere in the loop — choosing targets, approving payloads, or triage. GTIG's finding removes that bottleneck: vulnerability discovery, exploit development, and execution can now run as one automated pipeline against real victims.
That changes the economics of offensive security. Low-complexity flaws that were previously not worth a human attacker's time become viable targets when an agent can churn through them continuously. It also expands the attack surface in two directions at once: agents themselves are becoming targets, through the configuration files and cloud credentials they depend on, while the model access powering them is being stolen and resold at industrial scale. For defenders, the takeaway from Google's report is that the race is no longer theoretical — autonomous offensive tooling has moved from research demos into observed intrusions.
- #ai-security
- #vulnerabilities
- #threat-intelligence
- #cybercrime