Sign InOpen Brain
AI EngineerVideoSource Linked

We let an AI agent execute Bash and lived to talk about it — Sarah Sanders, PostHog

PostHog treats every context source as part of an agent’s supply chain, scanning at build and use time while reserving blocking decisions for deterministic controls.

AI Engineer · Sep 14, 2026
Open Source Open MarkdownOpen JSON
Source Summary

PostHog’s setup agent runs for about **8,000 users per week** and consumes docs, prompts and example apps as skill bundles. Its threat model includes poisoned first-party content, so inputs are scanned when skills are built and again when the agent uses them.

Practical Implication

For any agent that can execute commands, make Bash **deny by default**, keep secrets outside model context, and scan both incoming context and generated output. Enforcement should remain deterministic; an LLM may triage noise only after mechanical rules have decided not to block.

Agent-Ready Context
PostHog’s setup agent runs for about **8,000 users per week** and consumes docs, prompts and example apps as skill bundles. Its threat model includes poisoned first-party content, so inputs are scanned when skills are built and again when the agent uses them.

For any agent that can execute commands, make Bash **deny by default**, keep secrets outside model context, and scan both incoming context and generated output. Enforcement should remain deterministic; an LLM may triage noise only after mechanical rules have decided not to block.

PostHog reports almost no malicious prompt injection found in the wild and many false positives. Rule quality therefore depends on positive and negative tests, impact-based severity, telemetry and layered controls; no individual scanner or sandbox is sufficient.
Connected Context · Feed7 Judgment

This turns general warnings about Bash-capable agents into a deployed layered-control pattern and adds a subtle calibration lesson: real attacks were rare while false positives were common. It reinforces external, deterministic enforcement and least privilege, but narrows scanner-centric security claims because safe operation still depends on tested rules, secret isolation, telemetry, sandboxing and bounded authority rather than any single detection layer.

IT Admin for the AI Workforce — Sarthak Aggarwal, DecaworkDecawork extends deny-by-default Bash into identity and authorization: scoped capabilities, revocation and audit receipts must be enforced outside the model for every action.Give the Agent a Budget, Not a Token — Sachin Malhotra, AnthropicAnthropic’s deletion incident shows why an allowed command still needs quantitative blast-radius limits, rate controls and human-held overrides.From coding to Knowledge work agents — Karan Vaidya, ComposioComposio generalizes the same control boundary beyond Bash through action records, enforced permissions and preflight checks for irreversible operations.apache/makaMaka supplies complementary runtime primitives—sandboxing, durable event history, replay and recovery—that make blocked or permitted agent actions inspectable and testable.
Context Map
agentcodingsecurity#harness-engineering#sandboxing#agent-reliability
Uncertainty
PostHog reports almost no malicious prompt injection found in the wild and many false positives. Rule quality therefore depends on positive and negative tests, impact-based severity, telemetry and layered controls; no individual scanner or sandbox is sufficient.