AI & Tech

Google says attackers used AI alone to find and exploit software flaws

Google's threat intelligence unit says a tool found on an exposed server ran the full cycle of finding and exploiting flaws with no human involved. Victim data points to French media, entertainment and education websites.

AI text based on reports by The Next Web, ZDNet, adevarul.ro and other publishers. Published . How it is made

For membersThe first 30 seconds for everyone. Synthetic voice (AI).

In brief

How it came to this
  1. February 2026Best available models fail to find the bug after 25 hours and $350 of scanning
  2. July 2026GPT 5.6 Sol finds the flaw in 44 minutes for $22
  3. September 2026GPT 6 Sol finds the flaw in 13 minutes for $4.24 on its first guided prompt

An autonomous tool and its targets

Google's Threat Intelligence Group and Mandiant said on Monday, 16 November, that attackers had used AI to find software flaws and exploit them with no human involved, something the group said it had not seen before. The finding appears in the group's latest AI threat report. Investigators located the tool, which they call exploit_agent, on an exposed staging server. It works with several frontier and locally run AI models, compares popular web software before and after each patch, flags likely flaws, and then writes and runs exploits on its own. The previous quarterly edition, issued two months earlier, had said the group had not yet seen fully autonomous attack pipelines used against targets.

The same server held data stolen from victims, along with scripts for brute-forcing passwords, installing webshells and removing webshells planted by rival attackers. A configuration file carried comments in Russian. Based on victim data, the operator appears to have targeted French media, entertainment and education websites. The report does not establish that the operators acted on the orders of Russian authorities. Google described the shift in direct terms.

“fully automate vulnerability discovery and exploitation without a human in the loop for the first time”

Google

From patch to exploit in days

Google argues that patched flaws are easier for AI to attack than unknown ones, because attackers have a limited body of material to study: code before and after changes, release notes and published research. The report points to two PaperCut print server flaws that attackers chained together in late August. Researchers and the security firm Huntress used AI to build working exploits within a day of disclosure. GreyNoise, as cited by Google, said an attacker likely used AI four days after disclosure to run a campaign against hundreds of organisations. After PHP flaws were patched in October, hundreds of proof-of-concept exploits appeared that did not work, and the report says AI likely wrote many of them.

Open-source frameworks and the FBI's Artex warning

Google said several groups now use open-source AI hacking frameworks in live attacks, including financially motivated gangs and suspected Chinese spy groups. It saw the CAIRN framework used against targets including the US aerospace sector, and it expects attackers to rely less on humans and become stealthier as the tools improve. In a flash cybersecurity advisory, the FBI warned that hackers were using an open-source tool called Artex against financial services, government, tech, defense, phone and internet companies, and law firms. The agency said Artex intrusions can be noisy, yet the speed they allow makes it far harder for human defenders to keep up. In some cases Artex-linked attacks go from targeting a system to active compromise 'in under three hours'. Artex has been used since its first public debut in July, including against South Korean financial institutions where at least 68,000 people had their personal data stolen. The FBI also confirmed earlier Microsoft findings that hackers used Artex against ConnectWise and reached the Microsoft 365 cloud environments of at least 2,250 of its customers, and Microsoft says the hackers are likely Chinese.

Cheaper access and cheaper zero-day discovery

Google's latest tracker found criminals reselling stolen enterprise AI access through 'LLM Mixers', underground storefronts where AI token batches sell for as little as $0.10 while the organisations whose accounts were hijacked face usage bills above $100,000. The report describes LLMjacking as the theft of AI credentials such as usernames, passwords and API keys. John Hultquist, chief analyst of the Google Threat Intelligence Group, said the group had tracked LLMjacking rising through 2026. Separately, Xint.io tested seven models launched since February 2026, two from OpenAI, three from Anthropic and two open-source models from DeepSeek and Z.ai, on a flaw in Google's V8 engine tracked as CVE-2026-102321, a type confusion bug that was absent from the models' training data. In February the best models then available failed after 25 hours and $350 of scanning. GPT 5.6 Sol found the flaw in July in 44 minutes for $22, and GPT 6 Sol found it in September in 13 minutes for $4.24 on its first guided prompt. Only three of the seven models found the bug: GPT 6 Sol, GPT 5.6 Sol and Claude Opus 5.

Cost to find the V8 bug by modelUSD
GPT 5.6 Sol (July 2026)
22 USD
GPT 6 Sol (September 2026)
4.24 USD

Data from the article text

Wider pressure on defenders

Governments and critical infrastructure operators are still dealing with serious flaws found in PHP, a language that underpins about 70% of the web. Authorities in Europe are also facing attacks on supermarket refrigeration systems, some of which have been partly blamed on Iran. The new development adds to concern about AI-driven attacks, a worry that has grown since Anthropic introduced its Mythos model in April. Google expects attackers to rely less on humans as the tools improve.

Ask about this text

Sources and context

Organizations
Google, Anthropic, OpenAI

Pollar is free and carries no ads in the text. Its members keep it going.

Support Pollar