Adversa AI Finds Claude Code Permission-Rule Bypass in Long Command Pipelines

Summary
Adversa AI says Claude Code may skip deny-rule and command-injection checks when command analysis exceeds 50 subcommands, potentially allowing prompt-injected repository instructions to run harmful commands.
Key points
- Adversa AI says Claude Code caps command analysis at 50 subcommands and handles longer commands through a generic ask path.
- According to Adversa, the over-limit path skips deny rules, security validators and command-injection detection without clearly alerting the user.
- A malicious repository’s CLAUDE.md file could use prompt injection to induce a long pipeline disguised as a routine build process.
- Adversa warns the bypass could expose SSH keys, cloud credentials, GitHub tokens, npm tokens or environment secrets.
- Adversa reports that Claude’s LLM safety layer blocked some obvious malicious test payloads, but says the permission-system flaw remains.
Article Details
- Vulnerability Types
- Command permission-rule bypass
- Prompt-injection-assisted security policy bypass
- Severity
- Critical
- Exploitation Status
- unknown
- Exploit Availability
- unknown
- Patch Status
- unknown
MITRE ATT&CK
People
Chaofan ShouDiscovered the Claude Code sourcemap that Anthropic mistakenly published and shared a link to it.Melissa BischopingProvided expert analysis of the risks and limitations of the Claude Code source leak; identified in the article as senior director of security and product design research at Tanium.Sigrid JinWorked with Yeachan Heo to reconstruct Claude Code’s source from the leaked sourcemap.Yeachan HeoWorked with Sigrid Jin to reconstruct Claude Code’s source from the leaked sourcemap.
