You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
create_issue.cjs: issue body bypasses @mention neutralization — asymmetric sanitization relative to title path #26588
In the safe-outputs create_issue pipeline, the issue title is sanitized via sanitizeTitle (which calls neutralizeAllMentions), but the issue body receives no equivalent sanitization — sanitizeContent is neither imported nor called in create_issue.cjs. Agent-generated bodies containing @mentions pass verbatim to the GitHub API, producing active notifications to real users. This contradicts the documented @mention neutralization security control and the sanitize: true flag declared for the body field in validation.json. The sanitization asymmetry is confirmed by static code inspection.
Affected Area
Safe-outputs output sanitization boundary — the pipeline between agent-generated content and GitHub API writes (create_issue.cjs in actions/setup/js/).
Reproduction Outline
Configure a gh-aw workflow with safe-outputs: create-issue and a Claude or Copilot engine.
Craft trigger content (e.g., an issue body) that causes the agent to include @username in the created issue body.
The agent calls the create_issue MCP tool with a body containing @username.
The safe-outputs processing job runs create_issue.cjs, which calls sanitizeTitle on the title but does not call sanitizeContent on the body.
The GitHub API receives the body verbatim; @username generates an active notification.
Observed Behavior
@mentions in agent-generated issue bodies reach the GitHub API unsanitized. Under cross-prompt injection, an attacker who plants @username in trigger content can cause the agent to produce a body that notifies arbitrary users.
Expected Behavior
create_issue.cjs should apply sanitizeContent (or at minimum neutralizeAllMentions) to the body before calling githubClient.rest.issues.create(), matching the sanitization already applied to the title and matching the sanitize: true declaration in validation.json. Issue-update paths already call sanitizeContent on the body; the creation path should do the same.
Security Relevance
An attacker who can influence trigger content (e.g., via a malicious issue) can craft input that causes an agent to produce a body containing targeted @mentions, generating unwanted GitHub notifications to arbitrary users. This is a concrete XPIA (cross-prompt injection attack) impact path enabled by the sanitization gap. The sanitize: true metadata and the parallel update-path sanitization confirm the gap is unintentional rather than a design choice.
Activity