Blog / cluster
Runtime defence for LLM agents: prompt injection, indirect PI, tool-call gating, taint propagation.
The load-bearing question in agent safety is not whether the LLM will hallucinate — it will — but whether that hallucination can trigger side effects before a human catches it. On the 5,525-record ToolGuard-Bench (arXiv:2603.12621), rules-only defences block just 1.2% of malicious tool calls; the three-layer cascade shipping in AEGIS reaches 99.9% at 1.06 ms P50 latency. These articles cover the underlying attack surface and the defensive patterns that hold up in production.
2026-06-29
Indirect prompt injection hides instructions in webpages, emails, and tool outputs that your agent later reads. Five attacks with concrete defenses.
Load-bearing stat: IPIGuard (EMNLP 2025 oral) drops attack success rate from 4.43% → 0.69% by tracking parameter-level data flow rather than blocking by classifier.
Read →2026-06-29
Most teams log agent decisions to Postgres and call it done. Here's the 30-minute setup for production-grade tool-call auditing with structured events, retention, and tamper-evidence.
Load-bearing stat: Median time to implement minimal tool-call auditing in AEGIS: 27 minutes (docker compose + 2 env vars). Median time to satisfy a SOC 2 auditor: 1 day.
Read →2026-06-29
LangChain agents are particularly exposed to prompt injection because of tool chaining and retrieval-augmented context. Here's the defense stack that works.
Load-bearing stat: Tool-call gateway defense reduces LangChain agent attack success rate from ~12% (prompt-only mitigations) to <1% (gateway + parameter taint).
Read →2026-06-29
Runtime safety catches unsafe agent actions in real time, between decision and tool execution. What it covers, how it works, why it differs from LLM content safety.
Load-bearing stat: Of attacks that lead to actual harm in agent deployments, 86% happen at the tool-call boundary — not in the LLM prompt. Runtime safety is where the harm gets stopped.
Read →