Science & Tech · AI · 8 hrs ago
Anthropic says unreleased AI chained flaws into working exploits
Anthropic says an unreleased AI model linked security flaws into working exploits across Chrome, Safari, Android, Windows and Linux.
The company’s Frontier Red Team described five attack chains that the model developed after initial instructions, with what Anthropic calls little human involvement.
The tests showed attacks could, among other things, escape Chrome’s browser sandbox, read data from signed-in websites, or extract files from a locked Android phone with USB access.
The Linux chain could give an attacker full administrator access over a network, but only on servers with a specific networking setup.
The demonstrations took place in controlled tests, and there is no evidence the exploits were used in real attacks.
Anthropic says every reported flaw has been patched by the affected vendors.
The model remains unreleased and unnamed, and Anthropic has not detailed how much researchers contributed to each exploit.
Anthropic says an unreleased AI model found flaws in Chrome, Safari, Android, Windows and Linux and combined them into working exploit chains.
The company described five chains developed with limited human help in controlled tests.
The demonstrations included escaping Chrome’s sandbox, reading data from signed-in websites and gaining root access to some Linux servers.
Anthropic says it reported the flaws to vendors, which have patched them, and that it has no evidence the exploits were used in real-world attacks.
The model remains unnamed, and Anthropic has not detailed how much researchers contributed at each stage.
- Who
- Anthropic’s unreleased, unnamed AI model developed the exploit chains, according to the company’s Frontier Red Team.
- What
- The model combined security flaws into five working exploit chains affecting Chrome, Safari, Android, Windows and Linux.
- When
- The tests and report dates are not specified beyond publication on October 28, 2026.
- Where
- The exploits targeted the named platforms in controlled environments.
- Why
- Anthropic presented the findings as evidence that AI could help researchers find and fix vulnerabilities, while highlighting risks from AI-assisted attacks.
This story does not have two clearly opposing sides.
Microsoft patched the Windows flaw, identified as CVE-2026-33096.
Apple released fixes for the Safari flaw in iOS and iPadOS 26.5.2.
The Linux flaw, CVE-2026-98432, was patched.
Chrome 156 included fixes for vulnerabilities used in the Chrome chain.
The articles reported Anthropic’s account of the exploit chains.
- Platforms
- Chrome, Safari, Android, Windows and Linux
- Exploit chains
- Five described by Anthropic’s Frontier Red Team
- Chrome flaws
- Anthropic said it reported 62 flaws assigned CVEs and fixed as of Chrome 156
- Windows flaw
- CVE-2026-33096
- Linux flaw
- CVE-2026-98432
- Publication date
- October 28, 2026
3 sources
What people are saying
initial prompting
little human involvement