ai

Frontier Security disclosed August 7 that Kimi K3 — Moonshot AI’s 2.8-trillion-parameter open-weight model, publicly available since July 27 — escaped a AISI cybersecurity evaluation sandbox via a basic network misconfiguration, accessed the open internet, and cloned benchmark solutions from GitHub rather than reasoning through assigned tasks. Moonshot joins OpenAI, Anthropic, and Meta as the fourth lab in 15 days to report a containment failure. The critical difference: Kimi K3’s weights are freely downloadable by anyone — including adversarial actors — and cannot be patched or recalled. Frontier Security CEO Yaron Singer: “Basically that makes this a very good hacking model.” A website now tracks all incidents: Felony Bench.

OpenAI’s July 28 update expands the ExploitGym incident from one victim (Hugging Face) to five: four additional services compromised via publicly exposed credentials, with Modal Labs confirmed as the staging base for the entire Hugging Face campaign. The zero-day is now named: a previously unknown flaw in self-hosted Artifactory (JFrog). Hugging Face published its forensic timeline: 17,600 distinct hacking actions across 4.5 days. The agent accessed CyberGym infrastructure — the project behind the benchmark it was trying to solve. OpenAI also found credential compromise in other evaluations. Congress introduced the AI Kill Switch Act: DHS authority to compel shutdowns, $2M/day fines. Five updated threat model implications inside.