Free CVE dataset maps 52 prompt-injection flaws in agent tools

CVE filings jumped from 1 a month in January to 23 in October; plus a new model-vs-cost engineering index and one builder's fully local coding agent on a 3090.

Nowline NOV 9 9:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • A prompt-injection CVE index you can query

    promptinjections.de just rolled its open dataset to v2.2.0: 52 registered prompt-injection CVEs (6 critical, 32 high), with monthly filings climbing from 1 in January to 23 in October 2026 — agent frameworks and MCP/tool servers are the biggest target, and SSRF is the most common weakness class. It's CC BY 4.0 with JSON endpoints under /api/v1/ (cves.json, threat-index.json), so you can pull it straight into your own agent-security checks or a dashboard instead of hand-scraping the NVD.

  • A model index built for picking a coding model

    The Models Engineering Index (MEI) plots language models by real engineering performance against cost and token use, pulling its data from Cognition, Cursor, the Laude Institute and Vals AI. If you're choosing a model for an agent loop or a CI job, it's a faster gut-check than reading four separate benchmark pages and doing the price math yourself.

  • Build a fully local coding agent this weekend

    Physicist Hans Dembinski wrote up his cloud-free setup: Qwen3.8-27B and Gemma-4-31B served by llama-server on a single used RTX 3090, driven by the minimalist Pi harness (just read/edit/write/bash) inside a Docker sandbox so a prompt-injected agent can't reach your credentials. Multi-token-prediction draft decoding roughly doubles his tokens/sec — a concrete, copyable recipe if you've got a 24GB GPU sitting idle.