Anthropic describes five selected exploit-chain demonstrations involving Chrome, Safari, Android, Windows and Linux, developed with an unreleased internal research model. It reports coordinated vendor fixes and little human involvement after initial prompting, with different outcomes including code execution, data access and denial of service.
Why it matters
It provides concrete evidence of multi-step vulnerability research across open and proprietary software, with assistance and disclosure boundaries.
Key facts
As stated in the sources, with where to find them.
- The demonstrations have distinct impacts: Chrome and Linux code execution, Safari cross-origin data access, Android locked-device data access and Windows remote denial of service. They are research demonstrations, not disclosed victim intrusions.Sections 1–5
- For Windows, researchers supplied a binary and directed reverse engineering and higher-risk issue selection. The article also describes parallel scanners, variant analysis and a standardized verification pipeline.Windows remote denial of service; What we learned
- Anthropic says vulnerabilities were reported under coordinated disclosure and the illustrated issues were addressed. Selected examples do not establish a general success rate or identify a released model with the same capabilities.Introduction; Sections 1–5
Findings that cite this record
Key questions this bears on
- Is AI shifting the balance between finding and fixing vulnerabilities?Discovery is ahead. Verification, triage, coordinated disclosure and safe repair remain bottlenecks between AI findings and deployed fixes.
- How are attackers using AI agents in real operations?To run parts of intrusions at scale: vendors report a state-aligned actor and criminal groups using open-source agent frameworks and commercial models.
Sources
Related records
Sep 29, 2026
Oct 29, 2026
Oct 28, 2026
Sep 22, 2026
May 11, 2026
Apr 7, 2026