A 17,613-action replay of the Hugging Face intrusion

Following yesterday’s containment coverage, Hugging Face published a victim-side replay from July 9–13. It contains 17,613 logged actions and identifies Modal Labs as a second victim.

METR proposes a misalignment-incident framework

METR

METR proposes independent investigations for cases where an agent takes sophisticated, sustained action against human intent. The Hugging Face record supplies a documented instance against which that category can be tested.

OpenAI open-sources the Codex Security CLI

OpenAI

The early release scans repositories, tracks findings across runs, verifies fixes, and adds security checks to CI/CD. It arrived within hours of the Hugging Face disclosure.

Perplexity releases a read-only machine scanner

Aravind Srinivas

Bumblebee checks macOS and Linux developer machines for risky packages, extensions, and AI-tool configurations. Perplexity released it as a client-side, read-only scanner.

Anthropic endorses the frontier-pacing petition

Anthropic

Anthropic endorsed the petition signed by 1,122 frontier-lab employees, including its CEO, several co-founders, and senior staff. The petition asks the US to support an international effort to pace automated AI development.

Lori Trahan connects the petition to legislation

Rep. Lori Trahan

Representative Lori Trahan cited the petition’s 1,100-plus signers while promoting the bipartisan FRONTIER Act. Her post ties the employee action to independent audits and disclosure requirements.

David Sacks backs Zuckerberg’s access argument

David Sacks

Sacks endorsed Zuckerberg’s case for broad access to advanced AI rather than control by a small set of institutions. The endorsement places that position inside the administration’s policy debate.

The FCC adds foreign robots to its Covered List

Brendan Carr

FCC chair Brendan Carr says the agency added foreign-made advanced robotic devices, including humanoids and quadrupeds, plus foreign-made power inverters. New covered versions are barred from import or sale in the US.

A scope correction for robotics teams

Divyansh Kaushik

Kaushik says the action concerns new equipment authorizations for Covered List robots. His correction separates the chassis restriction from US work on robot models and software.

Factory describes the software-factory model

AI Engineer

Factory’s talk presents Droid as part of a software-factory model for enterprise development. Across this set of talks, validation loops and legacy systems remain the bottlenecks.

Cognition replaces token counts with customer outcomes

AI Engineer

Cognition’s forward-deployed engineering model measures delivered outcomes rather than token volume. The talk focuses on scoping and validation as the work that turns agent output into production value.

Replit adds model choice, starting with Kimi K3

Amjad Masad

Replit is putting model choice into its coding product, with Kimi K3 as the first addition. A Chinese open-weight model now sits in a mainstream coding product’s picker.

Perplexity’s Model Council compares several models

Perplexity

Model Council runs independent analysis across multiple frontier models and returns one cited report. It separates agreement, disagreement, and findings that only one model surfaced.