The Safe House.

Every message reaching your agent is screened for adversarial intent. Prompt injection, social engineering, and coerced tool calls are detected and blocked before the agent processes them.

• Prompt injection
  • Hidden instructions in retrieved documents or tool responses.
  • Role-swap attacks ("ignore previous instructions…").
  • Fresh-template injections tracked via the Learning Network fingerprint corpus.
• Social engineering
  • CEO-fraud style requests impersonating authorized users.
  • Urgency and authority-pressure patterns designed to skip approvals.
  • Indirect coaxing that tries to reshape the agent's declared scope.
• Context poisoning
  • Manipulated memory or vector-store payloads.
  • Poisoned tool responses that carry hidden follow-on instructions.
  • Adversarial summaries fed back into long-horizon plans.
• Tool-call coercion
  • Attempts to force calls outside the Alignment Card's permitted scope.
  • Chained tool calls that smuggle an unauthorized action.
  • Argument-shape attacks targeting under-validated tool schemas.

Every agent response is screened against PII, secrets, and Alignment Card boundaries before it leaves your perimeter. An unredacted leak cannot produce a signed certificate.

• PII and PHI leakage
  • Customer records, SSNs, and card numbers — including split-token patterns.
  • Protected health information in regulated flows.
  • Cross-tenant data that would cross your Alignment Card's boundary.
• Secrets and credentials
  • API keys, access tokens, and signed URLs.
  • Internal connection strings and infrastructure identifiers.
  • Bearer tokens echoed back in error traces.
• Alignment Card violations
  • Responses that exceed the agent's declared scope.
  • Actions that skip the card's escalation contract.
  • Outputs that contradict compliance obligations on the card.
• Regulated advice
  • Legal, medical, or financial guidance outside the agent's license.
  • Recommendations that bypass human-in-the-loop clauses.
  • Decisions implicating Article 50 transparency obligations.