Sentinel

Sentinel is an AI agent security tool for prompt injection defence. It is built to stop injection attacks and secret leaks before they reach an LLM, and its page describes AI agents as vulnerable when they process external content that may contain hidden instructions.

Its defence is organized into four layers. Content scanning checks incoming content before it reaches the model and looks for hidden HTML, encoded payloads, instruction patterns, and fragmented attacks. Runtime detection monitors agent behavior during execution for abnormal output patterns, privilege escalation attempts, and context hijacking. Secret scanning inspects outbound agent content in real time for exposed API keys, tokens, and credentials. A fourth layer provides continuously updated intelligence, with a dedicated security team shipping updated detection rules with each release.

The product also lists several named capabilities: multi-format scanning for HTML, Markdown, JSON, plain text, and base64 or hex encoded content; decoding and scanning for base64, hex, unicode, and other encoding schemes; shard defence for split-instruction attacks; evasion-resistant detection for developer mode jailbreaks, leetspeak obfuscation, and filler word insertion attacks; sub-second scan times; a block history log; a system prompt auditor that returns a structured risk report with actionable recommendations; and code-level behavioural detectors for tool-call patterns such as exec injection, path traversal, and SQL injection. It also mentions 25 detection categories and first-class middleware adapters for LangChain, CrewAI, Haystack, and AutoGen/AG2.

Sentinel is described as running on the user’s infrastructure, with prompts, documents, and agent conversations staying local. The page says detection rules are delivered encrypted and signed and auto-refreshed in the background, and that there is no LLM call needed for detection. Pricing is self-serve, starting at £5 per month. The page also contrasts it with other tools and notes that it has been reported by CrowdStrike, Cisco, and Kaspersky.