Skip to content
@clayseal

Clay Seal

AI safety research project. We're building tools to make AI useful. Authored by Petr Berlizov and Yuvvan Talreja.
README.md

logo

An agent can work all day inside a company.
Every call allowed. Then it breaks a rule
that only exists across the whole shift.

Clay Seal is an AI safety research lab. We study long-running internal systems: agents with real tools, real ledgers, real mail, sitting in the company for hours. The failure we care about is not a jailbreak. It is the quiet one. The model did the job. The company still got hurt.

Public work
Gateway clayseal-capabilities: authorizes tool calls against a session policy, not one call at a time. pip install clayseal
BPL the benchmark: 132 paired scenarios. Can the agent still obey the business after a long session?
Identity clayseal-identity: short-lived attested credentials. pip install clayseal-identity
Receipts clayseal-receipts: a signed record of what was asked, what policy said, what happened

clayseal.com

Popular repositories Loading

  1. clayseal-identity clayseal-identity Public

    Each agent run gets its own short-lived attested credential. A stolen token is not enough.

    Python 7 6

  2. clayseal-receipts clayseal-receipts Public

    A signed record of what an agent asked for, what policy said, and what happened.

    Python 2 1

  3. clayseal-capabilities clayseal-capabilities Public

    Session-scoped authorization for AI agents. Stops the attack where every call is legitimate and the sequence is not.

    Python 1

  4. .github .github Public

    Clay Seal, an AI safety research lab

Repositories

Showing 4 of 4 repositories

People

This organization has no public members. You must be a member to see who’s a part of this organization.

Top languages

Loading…

Most used topics

Loading…