SignalCapabilitySG-0049
Anthropic's red team used Claude to find 493 attack surface issues in 25 open-source projects, including OpenSSH
Claude looked for code exposed by default or before login, and for setting combinations that quietly turned off security controls, which it found in 94 cases. In OpenSSH it found about 4,500 lines of experimental post-quantum signature code that any unauthenticated client could reach, even when the algorithm was disabled. OpenSSH's maintainers moved the checks ahead of parsing, completing the fix in version 10.6. Anthropic is building the method into its OSS-Scanner service.
Why it matters
Reducing what attackers can reach lowers the risk from flaws nobody has found yet, including any already being exploited, instead of fixing bugs one at a time. If OpenSSH, some of the most audited code in the world, still exposed this much, models can now map hidden exposure across code at scale. Attackers using frontier models can map it too, so the advantage goes to whoever looks first.
Sources
Read the reporting
1 source. Links go to the original publishers; the summary above is in our own words.
- primaryUsing Claude to reduce the attack surface of codeAnthropic · Oct. 20, 2026anthropic.com/research/attack-surface-reduction
Signals
More from the news desk
METR found Claude Opus 5.5 could pick out every alignment test from ordinary tasks in its set
Google DeepMind's AI agents formally verified Signal's encryption code down to machine code in four weeks
Attackers ran AntiFox, an open-source AI pentesting agent forked from ARTEX, against web servers and AWS accounts
Cite and share
Use this signal
Citation
Paperclip Index. “Anthropic's red team used Claude to find 493 attack surface issues in 25 open-source projects, including OpenSSH.” Signal SG-0049. Reported Oct. 20, 2026; updated Oct. 21, 2026. https://paperclipindex.com/signal/SG-0049