Aletheia Aegis 0.3.0 released: AI output guardrails
Understand this story
Vikas Shivpuriya of Aletheia AI Safety Research Core has publicly released Aletheia Aegis v0.3.0, a guardrail tool that uses deterministic rules to check AI agent outputs on CPU. The release information states a p95 latency of 0.019 to 0.029 milliseconds and single-core throughput exceeding 73,080 evaluations per second; credential leakage and life-safety issues are blocked outright (BLOCK), while other issues are marked as FLAG. The tool runs offline by default and does not retain response text; semantic judgment is optional and is pre-redacted, S4 remains an experimental feature, and the code is released under the Apache 2.0 license. The above latency and throughput figures come from the publisher's claims.
AI overview based on source reports · Updated 17 hr ago
Coverage timeline
Follow the coverage to explore different perspectives.
- Hacker News AIAletheia Aegis: Deterministic sub-millisecond output guard for AI agents
Aletheia Aegis v0.3.0 is an Apache 2.0 TypeScript/Node.js guard that inspects model output before it reaches users, running deterministic pattern and structural checks at a reported p95 of 0.019–0.029 ms and hard-blocking credential leaks and life-safety violations, with an optional pre-redacted semantic judge for other cases. It runs offline on CPU by default with zero retention of response text, and its nine failure signatures are split across the output plane (Aegis) and the action plane (aletheia-mcp). The author reports 215/215 bundled eval cases passing and a 90.2% held-out catch rate with 1.4% false alarms under the balanced policy with the judge enabled.
Activity for this story
- Current comparable activity
- 6
- Peak comparable activity
- 10Nov 21 15:00
- Change in the past 24 hours
- –
Trends compare the same fully observed participants, which may cover fewer participants than current activity. Move the pointer or tap the chart to inspect hourly activity; use the left and right arrow keys with a keyboard.