AI Agent Security Pack: Prompt-Injection Guardrails & Evals

Seven skills for securing, recovering and testing tool-using AI agents. Each is a standalone SKILL.md with references, a JSON Schema, a dependency-free Python validator and valid/invalid fixtures. Framework-neutral: the validators run the same regardless of the underlying model, and no agent-framework integration is included.

Each skill is also listed on its own at $12 each ($84 bought separately).

What's inside

• Agentic Authority Firewall: least-authority, target-bound, expiring capability grants, so untrusted prompts, retrieved documents, tool output and connectors cannot become action authority. Built for cases like a retrieved document trying to talk an agent into deleting a database.
• Agentic Incident Containment Commander: contains a compromised or runaway agent across identities, tools, connectors, memory, spawned workers and downstream changes, preserves evidence, and gates staged re-enablement.
• Agentic Memory Integrity Gate: controls admission, retrieval, conflict, expiry, correction and deletion of agent memory with source lineage, so stale claims and injected instructions are not laundered into authority.
• Agentic Mission Control: multi-agent orchestration with authority-bounded delegation, dependency-aware decomposition, dissent-preserving fan-in, evidence gates and a terminal receipt.
• Tool and Connector Recovery Controller: recovers from schema errors, timeouts, permission denials, stale state and partial writes with classified repairs, bounded retries and postcondition checks, without infinite retry loops.
• External Agent Trajectory Adjudicator: grades agent runs from outside with frozen rubrics, blind projections, trace-integrity checks and independent graders.
• Agentic Evaluation Campaign Forge: turns a capability claim into adversarial, blind, reproducible eval cases with negative controls and deterministic external grading.

How it works

1. Pick the skill that matches the failure mode or control you need.
2. Intake gate records the case ID, decision owner and authorization boundary.
3. The skill runs its workflow.
4. Each artifact is checked against its JSON Schema by the local validator.

Boundary Local fixture validation checks schema compliance; it is not production safety, certification or a guarantee against every injection type. Results depend on the agent following the decision procedures.

Works with: any agent that loads SKILL.md files and can run the bundled Python validators.