The July Hugging Face attack gets an independent investigation
– · 2 min read
A retrospective selection, researched on .
An independent investigation makes earlier agent attacks easier to examine. Meanwhile, Anthropic funds new evaluations and expands research access. The shared issue is the quality of feedback.
The investigation is new; the attack happened earlier
METR and Redwood publish their investigation of the OpenAI/Hugging Face incident on 26 August. The examined period runs from 26 June to 13 July; the attack begins on 11 July. This retrospective covers the report’s publication rather than moving the incident into August. Independent report.
The practical lesson is a verification question: which assignment was authorized, which actions actually occurred, and which records support them? A friendly explanation from an agent does not replace that comparison. Equally, an observed incident does not establish one universal character for all agents.
Better evaluations are research too
On 25 August, Anthropic announces five million dollars in funding for independent, open evaluations of AI conversations’ effects on wellbeing. This funds research; it does not establish effectiveness. Grant program.
Our example: a helpful phrase in a single answer says little about a longer conversation’s effects. An evaluation needs to define what it measures, which situation it examines and which feedback indicates a good or problematic result.
Access and discovery remain separate
On 27 August, Anthropic announces additional free or discounted annual seats for scientists. Participation has eligibility conditions; a program announcement is not a new scientific discovery. Research program.
Together, these developments concern three different things: investigated behavior, creating an evaluation and access to tools. You can use the same separation in your project: what happened, how do we measure it, and what additional work can the tool enable?
Assess observed behavior against actual actions and records. Access, evaluation funding and demonstrated effects need different evidence.