Security by Design17 independent layers
Each layer protects independently. If one fails, the others continue.
Explore Security Deep DiveOur philosophy
We don't claim the system is infallible. We assume it will be tested, pushed to its limits, and that one day a layer will break. That's why there are fourteen.
Every incident is documented publicly. Every flaw discovered strengthens the system. Transparency is not a risk — it's our best protection.
Click on a layer to see the details
Full transparency on what happened and how we responded.
PIN leaked via messaging
Cause
The security PIN was exposed in a message processed by a sub-agent. The old code was compromised through reasoning visible in session logs.
Resolution
PIN changed immediately. Client messaging architecture redesigned: dedicated monitoring daemon, isolated one-shot sub-agents. PIN no longer transits through messaging sessions.
Corrupted default agent
Cause
A secondary agent became the default agent instead of the main agent. For 3 days, all commands were routed to an agent without the right permissions.
Resolution
Default agent restored. Added a healthcheck that verifies default agent configuration at startup. Consistency alerts added.
Internal reasoning leak
Cause
The agent's internal reasoning (chain of thought) was included in a response, exposing implementation details.
Resolution
Output filter added to strip reasoning blocks before sending. Response format validation made mandatory.
The three new EasyClaw v2 security layers are now directly integrated into the defense stack above (layers 02, 05 and 17). They cover three distinct moments in an action's lifecycle:
Risk Classification
The permission classifier evaluates risk before the agent even attempts the action.
See all EasyClaw v2 featuresHook Interception
PreToolUse hooks intercept and can block the action at execution time.
See all EasyClaw v2 featuresAdversarial Verification
The verification agent validates results by applying an adversarial process.
See all EasyClaw v2 featuresTransparency on current limitations.
- Does not make autonomous financial decisions
- Does not contact persons not on the allowlist
- Does not modify its own security rules
- Does not retain sensitive data between sessions
- Does not self-replicate or deploy new agents without approval
