postmortem-writing — agentic threat model

Not certain from the listing — assumes standard LLM vulnerabilities where prompt injection could force the model to generate biased, non-blameless, or completely fabricated root-cause analyses, or leak system details from its context window.

Not certain from the listing — the skill processes incident timelines and logs. If these inputs are retrieved via RAG or external databases, there is a risk of data exfiltration or log poisoning to manipulate the postmortem outcome.

The skill injects structure and facilitation guidance into the host agent. If the orchestration framework does not properly sanitize the generated postmortem markdown or attempts to automatically execute the generated 'actionable follow-ups' without human-in-the-loop approval, it could lead to unauthorized system changes.

Not certain from the listing — depends entirely on the host agent's deployment environment. If the generated postmortem reports are compiled using local PDF/HTML generators, vulnerabilities in those compilers could be exploited via malicious incident data.

Not certain from the listing — there are no mentioned guardrails or evaluation mechanisms to verify the accuracy of the generated timeline or to detect hallucinated root causes, which could lead to incorrect engineering decisions.

Not certain from the listing — postmortems naturally contain highly sensitive data regarding system vulnerabilities and operational weaknesses. The listing does not indicate any built-in compliance controls, data redaction, or access policy enforcement for this sensitive output.

Not certain from the listing — if this skill is triggered automatically by other agents (e.g., an on-call alerting agent), a compromise in the upstream agent could allow an attacker to auto-generate misleading postmortems to cover their tracks during an active breach.