red-team-tactics — agentic threat model

Not certain from the listing — The underlying foundation model is not specified. Standard threats like prompt injection could bypass safety guardrails to generate actionable exploit payloads or bypass the intended read-only planning scope.

Not certain from the listing — The agent relies on MITRE ATT&CK mapping and tactics knowledge. If this knowledge base or its retrieval mechanism is poisoned, the agent could provide flawed, unsafe, or intentionally malicious red-team advice.

The agent framework restricts allowed-tools to Read/Glob/Grep via frontmatter. The primary threat is tool-use abuse, where an attacker uses prompt injection to coerce the agent into using Glob/Grep to locate and exfiltrate sensitive local files (e.g., credentials, source code).

Not certain from the listing — The hosting environment and sandboxing controls are not detailed. If the Read/Glob/Grep tools are executed without strict container-level path restrictions, the agent could access host-level configuration files.

Not certain from the listing — There is no mention of logging, guardrails, or evaluation frameworks to monitor the agent's output or detect anomalous tool execution patterns.

The agent implements a specific security control by restricting its tool scope to Read/Glob/Grep in its frontmatter configuration, preventing it from executing write or execution commands directly.

Not certain from the listing — While described as an 'Agent Skill' that could be integrated into larger multi-agent systems, specific ecosystem interactions or trust boundaries are not defined.