Four silent failures. Four fixes.
Automation fails quietly. These are the failures I found this year in production AI agents and detection pipelines, and what changed.
- FailedClaude Code safety rules were silently dropped past a 2,000-character cutoff.
- FixedRules fixed, plus an agent-monitoring playbook and recurring audits of agent instructions.
- FailedEnable calls reported success on disabled detection rules, and detections returned the wrong output columns.
- FixedLive canaries and tenant probes now expose both.
- FailedTwo production YARA rules, including Shai-Hulud v2 npm detection, were unmanaged.
- FixedDrift checks recovered both, and about 157 rules now sit under CI governance.
- FailedPhishing auto-close automation failed open.
- FixedRemediated, with AI instruction governance locked down.
On record
| 95% fewer | false positives after moving National Audubon Society to AI-driven detection, with 99.2% detection fidelity. |
|---|---|
| 157 rules | brought under version control and CI gates at a cloud-native fintech, including 81 builder rules and 13 YARA groups. |
| 55 to 60% lower | shared review cost from an in-house AI pull request reviewer, with fleet spend capped afterwards. |
| 13 years | in security, systems, and network engineering across fintech, nonprofit, enterprise, and industrial environments. |
Previously