Embedding models detect malicious AI agent sessions
Original title: Monitoring AI agent sessions: Leveraging embedding models to aid threat hunting and attack discovery workflows
Elastic stored Claude and Cursor sessions as vectors and used semantic search to flag one in seven suspicious sessions for triage.
Elastic Security Labs embedded six months of AI agent sessions as semantic vectors in Elastic and performed semantic search instead of LLM analysis on every session. The approach caught one in seven sessions requiring manual review of suspicious actions such as reading customer data exports and using Git to push to a remote repository. Unlike full LLM evaluation of each session, embedding-based search enables ongoing threat hunting on historical data without high per-session costs or loss of coverage.
- One in seven sessions flagged for triage
- Semantic search replaces keyword or regex rules
- Avoids reevaluating every session for new questions
- Applies to unstructured agent reasoning and tool calls
- Vendors
- Elastic
- AI models
- embedding models
From the original article
15 min read at elastic.co
I tested vector search on 7,454 turns of my own agent sessions. One in seven sessions needed review, the agent’s reasoning carried more signal than the prompt, and the stored vectors answer questions nobody thought to ask at ingest.
By Marcus Jeffes · elastic.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.