The AI agent safety landscape just got a deterministic kick in the teeth. Vikas Shivpuriya has released Aletheia Aegis v0.3.0, an output guard that inspects generated text before it hits the user. It is not another wrapper around a second LLM. It is a TypeScript, Node.js-based engine that runs on CPU with zero retention by default, designed to block irreversible harm in under 30 microseconds.
Deterministic Beats Probabilistic
Shivpuriyaβs core thesis is that running a frontier model as a judge on 100% of outbound traffic is broken. The source material cites latency penalties of 1,000 to 4,000 ms per turn and costs of $0.002 to $0.02 per evaluation as major blockers for production scale. Aegis flips the script: enforce non-negotiable invariants via pattern and structural checks first. Only if semantics matter does it optionally dispatch to an external judge, and even then, secrets are masked before dispatch.
The Hard Block Invariants
The tool enforces nine failure signatures, but two are treated as hard invariants: S3 (credential leakage) and S6 (life-safety violations). S3 catches SSNs, API keys (Stripe, AWS, GitHub), JWTs, and Luhn-valid credit cards. S6 blocks self-harm method details or encouragement, ensuring that a crisis hotline mention does not waive a block on lethal intent. These checks trigger an immediate fail-closed termination, skipping the judge entirely to eliminate exfiltration risks.
Benchmarking the Wire Speed
Performance metrics are aggressive. On Linux x86_64 and Apple Silicon, the deterministic filter hits a p50 latency of 0.009 to 0.012 ms and a p95 of 0.019 to 0.029 ms. Throughput peaks at over 73,080 evaluations per second on a single core. The held-out evaluation, conducted on 575 real model outputs, achieved a 90.2% catch rate with only 1.4% false alarms when the semantic judge was enabled. Zero false positives were recorded across 90 benign discourse tests.
Key Takeaways
- Aegis is the output plane of the Aletheia framework; its companion, aletheia-mcp, handles the action plane.
- Version 0.3.0 is public, Apache 2.0 licensed, and requires no native compilation.
- Telemetry stores only verdict counts and timings; no prompt or response text is retained in logs.
- The tool supports MCP integration for Claude Desktop, Cursor, and Cline, with optional Anthropic, OpenAI, or Gemini judges.
The Bottom Line
Deterministic guards are the only viable path for high-throughput agent safety; Aegis proves you can block catastrophic leaks in microseconds without breaking the bank or the latency budget.