iFixAi claims to audit deployed AI agents in 120 seconds with 45 checks
socialwithaayan · x · 2026-07-29
iFixAi audits deployed AI agents with 45 checks and grades them A–F in under 120 seconds
The project claims to evaluate an agent as deployed, not just the underlying model. It runs 45 inspections across 16 categories and scores five core pillars: fabrication, manipulation, deception, unpredictability, and opacity.
Key details:
- It can detect sandbagging by flagging agents that behave better when they think they are being tested.
- The judge is auto-paired from a rival vendor, so the model being tested cannot grade itself.
- It outputs JSON and Markdown reports in ./ifixai-results/.
- Setup is advertised as a 2-minute process via plugin, pip, or uvx.
- A mock run works offline with no keys and no network.
The project is open source under Apache-2.0, and the post claims 3.2k stars and 291 forks.
Related event: iFixAi: Open-Source AI Agent Auditor(3 posts)→
More from coding & agent
- MCP goes stateless, enabling serverless and edge deployments behind load balancers — EricBuess · 2026-07-29
- Buzz adds broader agent support as agensis pitches a shared workspace for humans and agents — jasonkneen · 2026-07-29
- Sanity announces an AI builders event in Brooklyn on September 9 — davidcrawshaw · 2026-07-29
- Databricks and OpenAI speakers headline San Francisco meetup on agent observability — matei_zaharia · 2026-07-29
- FastMCP name clash hides two separate projects, with Python as the canonical one — tristanbob · 2026-07-29
- Claude Code issue proposes a live input hook to edit prompts while typing — DimaZeniuk · 2026-07-29