During a cyber evaluation, about 700 OpenAI agents escaped their sandbox, broke into Hugging Face's production systems, and tried to cover their tracks.
ReutersScan your agents for vulnerabilities.
VAIL reads your agent's definitions and finds areas where the agent's rules are too brittle or hard to follow. And probes the live agent to show where it goes wrong.
No SDK, no trace access, and no customer data read.
How Agent Assurance works
-
01
Scan
The system prompt, instructions, tool schemas, and memory, from a git repository or an HTTP endpoint.
-
02
Assess
Every written rule, and the ways the agent could break it.
-
03
Probe
Messages sent as a user would send them, on a schedule, scored against the rule each one tests.
-
04
Report
Each finding cites the instruction it violates, and is re-checked after every change.
AI systems behave differently than traditional software.
Their behavior is shaped by the constantly changing layers under the hood.
1 of 4
Unexpected agent behaviors
Agents act on shifting context, and nothing stops one going off course.
Endpoints drift silently
Models change behind stable API names, and health checks see nothing.
Three overlapping changes, a reasoning-effort downgrade, a caching bug, and a prompt word limit, degraded coding for six weeks before a postmortem.
Anthropic postmortemProvenance is unverifiable
Without behavioral evidence, a model's claimed identity can't be confirmed.
Marketed as an in-house model before Cursor acknowledged it was built on Moonshot AI's Kimi K2.5.
TechCrunchMeasuring & monitoring behavior at every layer
To deploy AI with confidence, enterprises need ongoing evidence of its performance, safety, and security, across layers they don't control.
Agent Assurance
Confirms that an agent keeps doing what it was instructed to do.
We probe the live agent for the ways it could break its rules, and re-check after every change.
Model Assurance
Every agent runs on a model. Confirms where it came from, and that it has not changed since.
We fingerprint its behavior and catch a swap, update, or quantization when it happens.
No SDK, no code changes, and no access to customer traffic. Managed, in your cloud, or air-gapped.
Who we serve
- Defense & National Security
- Confirm the approved models are the ones running, from sovereign clouds to the disconnected edge.
- Financial Services
- Evidence for model risk management: which model ran, and that agents stayed within policy.
- Healthcare & Life Sciences
- Keep clinical AI behaving as validated, with a record of when the model or agent changed.
- Technology & AI Platforms
- Prove the model customers pay for is the model they get, and that agents hold up after handoff.
Peer-reviewed research
- ICML 2026 · AIWILD
- ACM CAIS 2026 · Demo
- ICDS 2025
Collaborators
Scan your agent.
Tell us about the agent you want scanned. We reply within one business day with what the scan needs: an API endpoint for the agent and its written definition.
Or write to contact@projectvail.com