Don’t Panic Digest Inform. Gate. Then automate. News — PraisonAI security advisories (GitHub, Oct 8): PraisonAI, an open-source framework for building SI (super intelligence) agents, published three advisories rated high to critical (CVE-2026-61426, CVSS 8.6; CVE-2026-61427, 7.3; CVE-2026-61445, 9.9). In its default setup, the agent server could be reached by anyone on the network with no login, and the tool-server mode had no login unless a key was set. A third flaw let a model’s tool calls write files and run commands without checks, so a prompt injection could lead to command execution. The advisories list fixes in releases 1.7.3 and 4.6.78, and newer releases are out. Not in CISA KEV; none of the sources we read reports exploitation. Operator read: upgrade, require a login before an agent server starts, keep it on a trusted network, and give a model’s tool calls only the file and command access a person has approved. https://github.com/advisories/GHSA-9mp3-24cc-77mg Related: UK AISI’s “Transect: Making large-scale agentic evaluations easier to understand” (Oct 7) — AISI says a final score reveals little of how a long agent run went; Transect, an open-source tool, lays the agent’s recorded steps out as one timeline, and AISI warns that SI-made labels “can be wrong or misleading”. For us, judge an agent by its record, and spot-check any SI summary of it. Our Plainwrap daily notes: https://x.com/CodeSolutionsIL/status/2108588340413100309 https://www.aisi.gov.uk/blog/transect-making-large-scale-agentic-evaluations-easier-to-understand More — Matt Turck × Andy Pavlo (agents and databases, ~04:09–10:51 + 15:10–30:43): What’s in it: Matt Turck and database researcher Andy Pavlo talk about what happens when SI agents get access to databases: copies (branches) for safe testing, agents that delete data, and why Pavlo says the old permission controls should apply to agents too. They also cover where agent memory should live and how well agents turn plain questions into database queries, with Pavlo’s numbers given as his own account. Our takeaway: Give agents the same scoped permissions you would give a person, let them test on a separate copy, and keep a named person’s OK on anything destructive. Notes of a public podcast; the speakers’ claims are as discussed, not verified. https://gist.github.com/CodeSolutionsLLC/6cc37a8e83a82d4689a3a3cd4b9f5319 https://x.com/mattturck/status/2108223135673696504 https://codesolutionsllc.com/news
