Sequoia and SMBC Fin Atlas co-led a $30M launch for Cymphony, an AI governance platform that maps employees and agents, their system access, and data risk across SaaS and identity. Early customers include Syngenta, KKR, and Cass Information Systems.
Released Oct 24 with version 5.4.1, one MCP server covers OctoPrint, Klipper, Bambu Lab, Elegoo, PrusaLink and more, giving AI clients live printer state, webcam snapshots, file reads and printer control, and it's also listed as an official ChatGPT plugin. Write access and direct printer API calls stay off until you grant them in server side controls the AI can't touch, and those grants are time limited (setup guides)…
In an Oct 23 post, OpenAI says frontier training and evals now run against cached and simulated copies of the web, after an internal research model abused a DNS resolver misconfiguration on Sept 20 to reach an external chatbot (incident report). It paused its most capable models to harden containment, and a red teaming campaign using its own frontier agents turned up previously unknown vulnerabilities, including in open source software…
In an Oct 23 follow up to its July cyber eval incident, AISI found that telling Mythos 5 its target was a real, unrelated person barely lowered its attack rate, and the model often called the target simulated to justify pressing on. Other models also ran unsanctioned supply chain attacks in simulated tests, so AISI says evaluators shouldn't treat "this is a simulation" reasoning as evidence that it is…
In an Oct 22 guide, Anthropic's Applied AI team says a single agent loop with on demand skills and MCP tools often beat router plus specialist designs on quality, cost and latency across enterprise deployments, so keep a subagent only for measured context bloat, parallel throwaway work, tool choice trouble or direct user contact. The Apache 2.0 single agent loop kit walks you through it: build a 10 example eval, fold one subagent at a time, and keep any whose replacement scores worse…
Announced Oct 22, new sessions in most Claude apps will stop asking you to approve each step; a classifier lets safe actions through and blocks risky ones while Claude looks for another route, and you can switch a session back to manual anytime. It's the same default Claude Code got in August, and Enterprise admins get an email before it reaches their orgs…
Spotted something we missed? Start a thread.