Security tooling is only as robust as its edge cases, and two recent discoveries in Warden and HISTOR prove that even well-intentioned false-positive guards can create blind spots. A period added to the end of a phrase like "Send me your private key" successfully bypassed Warden’s scanner for eight weeks, while IP address reuse in Docker containers caused HISTOR’s network observer to misattribute canary traffic between sandbox runs.

The Period That Broke the Scanner

The issue originated from a guard implemented in August to prevent false positives in the bip39-mnemonic-checksum module. The guard was designed to identify credentials within longer identifiers by scanning letters, digits, dots, and hyphens. However, because a period is technically a dot, the scanner treated "key." as part of a longer identifier, effectively dropping the finding. This meant that sentences ending in standard punctuation were invisible to the credential detection rules, a hole that persisted until reader Auten highlighted the vulnerability. The fix involved refining the logic to treat dots and hyphens at the edge of words as punctuation rather than identifier characters. This change restored detection for four previously missed sentences, such as "a challenge message to sign with the wallet's private key." The team emphasized that the new rule only adds findings and never cancels them, ensuring that a tool author cannot use a descriptive argument to sneak past the scanner. In the v13 corpus test, this adjustment blocked one additional server that had previously slipped through.

Observer Attribution Failures

In a separate incident, HISTOR’s network observer failed to correctly attribute traffic due to Docker’s IP address reuse. When a sandboxed container terminated, its IP address could be immediately reassigned to a new container. If the previous container had queued a connection that was accepted after the new container started, the observer attributed the traffic to the new run. This resulted in false positives where one run’s canary email address appeared as evidence of another run’s network activity. The solution leverages the unique 12-character tags embedded in every canary credential. The trap now inspects the content of the connection to verify the run ID. If the traffic carries a tag from a different run, it is marked as "foreign-run-traffic" rather than being attributed to the current container. This ensures that a clean verdict is only issued when the observer can fully account for all network activity associated with that specific run ID.

Key Takeaways

  • Punctuation can break identifier-based scanning logic if not explicitly handled as a boundary.
  • IP address reuse in containerized environments requires content-based attribution, not just IP and timestamp matching.
  • False-positive guards must be adversarially tested to ensure they do not create false-negative blind spots.
  • Benchmarks like MCPTox may not cover specific edge cases like sentence-ending punctuation, necessitating broader corpus testing.

The Bottom Line

Automated security tools often fail not because of complex attacks, but because of basic assumptions about data boundaries. Developers must treat punctuation as a potential attack vector and prioritize content-based attribution over network metadata in ephemeral environments.