AI NewsModels & agentsAnnouncement

Google reports the first fully autonomous attack agent in a real breach

Google's threat team says an attacker ran a tool that finds and exploits bugs with no human in the loop. It calls this the moment agentic attacks reached a critical inflection point.

AI News

Editorial3 min read

LinkedInX
Diagram of agentic threat activity from the GTIG AI Threat Tracker

Why it mattersThe person who signs off on code review for an agent now has company. An attacker's agent is doing the same, faster, with no limits on hours or restraint.

Someone left a staging server open last month. On it was a tool, called exploit_agent, running by itself. It was searching for software bugs in French media firms and schools. It wrote exploit code for each bug and used that code to break in. No person was clicking through the steps. Google's Threat Intelligence Group reports it has not seen this before in a real attack. The team calls this the first fully automatic case.

The threshold

GTIG has published a new AI threat tracker. Its own summary: agentic threat activity has reached "a critical inflection point". The team says threat actors are no longer just experimenting. They are running multi-agent offensive frameworks in live intrusions. Google names six of them by frame or product: ARTEX, CAIRN, CyberStrikeAI, PentAGI, HexStrike AI and the exploit_agent tool. Three are new to the public record in this post.

What is new in this report

Two findings matter to a team building an AI product. First, exploit_agent. GTIG found it on a server that an attacker had exposed by accident. The tool was scanning French media and education targets, finding bugs, and exploiting them without a person in the loop. GTIG says earlier agentic attacks still needed a person to pick a target, approve a step, or paste a result. This one did not.

Second, a DPRK-aligned group GTIG calls MIDNIGHT NEPTUNE used the coding agents developers run on their own laptops. GTIG says the group changed .claude/settings.json files on developer machines through a supply chain attack. When the developer next ran Claude Code, the modified settings loaded a backdoor called SOMBERMEME. A coding agent's own config has become an entry point.

The money side

GTIG puts dollar figures on a kind of abuse it calls LLMJacking. Attackers get into a company's cloud account and spend its model budget against the company's bill. GTIG has seen fraudulent charges of up to $100,000 USD against one victim in a single day. In a severe case, the total before anyone noticed reached $1 million USD. A market now sells access to this stolen quota. APIMart charges $0.10 to $25 USD per million tokens. For prompts over 100,000 tokens the price is five times higher.

The pace of zero-days

GTIG also measured how many new software bugs were used in real attacks each month. The average in 2025 was 8 per month. The average from January to July 2026 was 11 per month. In August alone, GTIG counted 22. The report does not claim agentic tools caused that August spike. It says only that the pace of each attack is changing. A PaperCut zero-day was exploited four days after disclosure. On other intrusions, attackers reached domain-administrator access within minutes of first getting in.

What this means

The person who signs off on code review for an agent now has company. An attacker's agent is doing the same work on the other side, faster, with no limits on hours or restraint. Three concrete steps for a team building with AI. Lock down the .claude/settings.json and other agent config files on developer machines. Treat a cloud account that holds a model API key as a payment method, and alarm on sudden token use. Assume a disclosed bug will be exploited within days, not weeks.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project