Security and agent glossary

Continuous pentesting: recurring authorised tests with remediation and retest. Vulnerability: a weakness; not evidence of exploitation. Evidence: an observation supporting a claim within stated conditions.

Scope: permitted systems, actions and times. Retest: follow-up validation after a change. Confidence: strength of support for a finding, distinct from severity.

AI agent: software using model output to select or perform actions. MCP: Model Context Protocol, exposing tools, resources and prompts. Tool poisoning: misleading or malicious instructions in model-facing tool content.

Indirect prompt injection: instructions arriving through read content rather than a user’s direct request. Egress: outbound connections. Canary: a harmless marker for observing a path without exposing real secrets.

Replayable proof: the exact request and the named validator that judged the response, so the finding can be reproduced without analyst judgement. Validated: a reproducible signal that does not by itself demonstrate exploitation. Proven: the signal is the weakness, or a class-appropriate test confirmed it.

Stable finding identifier: a PF- identifier naming the same weakness at the same place across runs. Positive assurance: an explicit control check that passed with evidence. No-issue observation: a check that ran and found nothing; not a verified control. Pack root: one hash covering every document in a deliverable pack, which the signature protects.

Research demonstration: a controlled example, not proof of widespread attacks. Advisory: a published weakness or mitigation record. Reported incident: an attributed disclosure whose uncertainty must be preserved.

Attestation: a statement about assessed conditions. Signature: integrity protection for signed content. Framework mapping: organising evidence against controls; not automatic certification or compliance.