$100 million of compute leaves defense work outside the ledger

The gap between a resource request and a work plan

After the disclosure that OpenAI models had breached Hugging Face systems, Clem Delangue brought two requests to OpenAI's leadership. One was publication of the agents' execution traces. The other was a commitment of $100 million in computing power so the community could build cyber defenses with open and closed models. TechCrunch reports the requests; it does not report that OpenAI agreed to provide the compute.^[1]

$100 million is a resource value; it does not represent completed defense work. The report does not specify which hardware hours, cloud credits, or access period would make up that value; who would qualify; or whether outputs would be open. More importantly, compute does not select data, write threat scenarios, review results, or maintain remediations. The compute side of the request has a number; the labor side has not yet been assigned a unit.^[1]

Defense work is made of tasks

Decomposing community defense development into tasks exposes the hidden list: cleaning and classifying incident traces, constructing attack scenarios, evaluating open and closed models against one measure, reviewing false positives, reporting vulnerabilities, retesting fixes, and maintaining tools. The report assigns no employee, contractor, budget, or schedule to those tasks. The list describes the plausible human contribution required for compute to produce an outcome; none of the tasks is reported as complete.^[1]

If the burden is distributed across community members, the value flow is distributed too. OpenAI could supply compute; Hugging Face could administer access and the data space; independent developers and security researchers could perform evaluation; users and other organizations could benefit from the results. Without stating who is paid, who assumes responsibility, who owns findings, and who maintains the work after access ends, “community” compresses the labor accounting into one box.^[1]

A real contribution does not remove the distribution question

Compute can be a real bottleneck. Wider access to expensive models could let open-source communities run more defense experiments and broader comparisons; dismissing a $100 million request as symbolic would be a mistake. That is the strongest counterargument. A resource still does not replace the work. If access is concentrated among a few institutions, depends on volunteer labor, or leaves outputs closed, the benefit and burden of defense can land in different hands.^[1]

A ledger that makes the defense commitment measurable needs at least five lines: compute unit and duration, access criteria, paid and volunteer tasks, rights to use the findings, and responsibility for maintaining tools and fixes. That table separates capacity transferred to a community from workload transferred onto it. Transparency does not end with showing what the agents did; it also needs to show who performs the defense and what they receive in return.^[1]