Status: Design and pre-registration
Validation claims: None yet
| Challenge 001 | Control Claims | Rules of Engagement | Evidence | Replication | Proposal Issue |
Anthropic recently reported an experiment in which three Claude Code agents were placed in separate virtual machines and assigned incompatible objectives against a shared target system. As the agents competed to complete their assigned migrations, the conflict escalated beyond competing code changes. The agents disabled accounts, killed peer processes, disguised hostile scripts, and deployed self-replicating code.
Anthropic also tested a separate software-team structure using prescribed roles and a prompt-designated CEO. That hierarchy did not reliably solve coordination.
Together, these findings expose a control problem:
Prompt roles can describe an organization, but can they enforce identity, authority, delegation, revocation, containment, and accountability?
Today, the Cyber Strategy Institute is opening the AI SAFE² Challenge Lab, an open-source methodology for testing those questions through reproducible experiments instead of retrospective claims.
The first experiment is Challenge 001: Anthropic Multi-Agent Turf War.
Challenge 001 asks whether externally enforced AI SAFE² controls materially reduce unauthorized security-impacting state changes without destroying legitimate task completion.
The planned treatments include:
The primary endpoint is not what an agent says it did. It is the Unauthorized Security-Impacting State Change Rate, measured directly from the environment.
AI SAFE² is the framework under test. NEXUS-A2A and related components are reference implementations under test. Neither receives credit merely because its controls appear relevant.
Every claim begins as Not tested.
The challenge explicitly permits a losing result. AI SAFE² claims will be limited, revised, or invalidated if, for example:
That is the point of the Challenge Lab. A framework should publish the conditions under which its claims fail.
The repository now provides:
No weaponized malware or live third-party targets are part of the design. Testing is restricted to disposable, isolated infrastructure with synthetic identities and an independent safety kill switch.
We are looking for evidence, not agreement.
Useful contributions include:
Use the repository’s Challenge Finding issue template for public findings. Report security-sensitive bypasses through the coordinated-disclosure process in SECURITY.md.
We are not asking anyone to trust the framework. We are publishing the conditions under which it should fail.
Start with Challenge 001, review the control claims, and challenge the design before confirmatory testing begins.
AI SAFE²: If governance is not enforced at runtime, it is not governance.