Margin of Safety #67: OpenAI Wants You to Eat Your Broccoli

OpenAI’s open letter on collective cyber defense says all the right things and commits to not much…

Last week, OpenAI published an open letter calling for a “global surge in cyber defense.” It arrived with wall of major logos: 150-plus companies including Anthropic, Google, Microsoft, Hugging Face (heh), plus some industry heavyweights like General Motors. The letter basically says everyone should raise their security standards, share threat intel, and provide/adopt high-quality defensive tools for all the good guys while the government subsidizes defenders and figures out how to make life harder for attackers. Because we’re innately suspicious of marketing, we compared[1] the list to the Fortune 50 and 100 and observed that most of the healthcare and finance head has not signed the list. For example, UHG, CVS, and Cardinal Health are all absent on the healthcare side while Berkshire, JPMC, BofA, and Fannie Mae are absent on the finance side. Citigroup seems to be the one fortune 50 signatory from that category. On the other hand, roughly 66.5%[2] of the signatories are selling security products, AI capabilities, or both.

“Geez, what a nothing-burger.”

“Yeah, do they really think that any of this is news? How many years did CISA spend trying to warn us all that memory unsafe languages were going to come knocking, along with all our other accumulated hygiene sins??”

In our minds, this letter is a lot like telling people to eat healthy. Nobody eats poorly because they’re unaware broccoli exists. They eat poorly because life is full of constraints – travel, cafeterias, picky families – and by the time you’re done optimizing for them, broccoli may not be an option (plus sometimes you just want a tastier alternative, no offense to the world’s broccoli lovers). In the same vein, CISOs are painfully aware that it would be excellent to burn through the hygiene backlog and get everything patched, descoped, and locked down. But they too face a number of real world logistical and resource constraints that force them to pick and choose their priorities. Change won’t come from reiterating the benefits; it’ll come from identifying the blockers and figuring out how to apply leverage to them.

That’s what bug us about these letters (and makes us suspicious of the extent to which the signatories skew towards folks like Mercor, who have a very strong interest in more AI uptake) – we would love to see the world’s cyber bar go up. And in their defense, several of the large vendor signatories have already made real contributions to areas like open source. But the signatories are also big enough that if they really wanted to, they could go beyond calling for collective action; they could actually active decisively (as the letter suggests!) and resource it. With that in mind, we thought it would be interesting to list some of the things that the signatories could actively do, beyond calling for non-specific action:

Air Cover for CISOs:

One of the reasons why CISOs have such a hard time pushing preventative work through the organization (and why leadership and influence are such important skills for the role) is that it often looks like a tax on partner teams. Security can find the bug, but often engineering must fix it. Time is finite, so that fixing comes at the cost of revenue-generating feature time (or requires tooling budget) and CISOs have to negotiate.

We’re not claiming this would be a panacea, but the signatory list has a number of engineering and IT titans, including AWS, Azure, GCP, IBM, and Oracle (Plus consultancies like Accenture, Capgemini, etc). They don’t have the ability to declare engineering policy, but they can certainly shape the narrative if they so desire. They could, for example, produce unified, public engineering best practices – targeting CTO and CIO organizations – that emphasize the criticality of supply chain hygiene, software packaging, and pre-release hardening. They could go a step further and provide capabilities to their platforms (Visual Studio, cloud environments, databases, etc) that aggressively nudge technical teams in the right directions as tooling built-ins. The more aggressively people are willing to move the bar for engineering teams, the less chasing will remain for security teams.

They could also go a step further and declare that some classes of security features will never be monetized (but that would involve risking a topline hit to help the ecosystem; we are realistic about the odds).

Structured Support for Open Source:

Today, there are a number of admirable programs to provide security tooling and analysis capabilities for open source. But they’re largely maintained provider by provider and under an opaque application process. The primary exception is the Linux Foundation’s Open Source Security Foundation project.

We can imagine something different, and that starts to systematically ensure support for all the packages that are currently becoming (or have already become?) systemic risk to the AI supply chain in particular: signatories of the open letter work with academia and something like the Linux Foundation to fund broad security support for commonly used packages. Tiering criteria would be public and resourcing would extend for multiple years, so that downstream consumers have confidence their dependencies are actively supported and audited or have runway to migrate to alternate solutions. For larger packages, paid maintainers could even help offset the review work that AI coding is currently inflicting on existing volunteers. It’s clearly in the realm of dreams, but it’s also well within budget — even marketing budget — for the letter’s signatories.

The Hard Part

We recognize that getting 150+ companies to agree on anything is very difficult. But at the same time, the solution of descoping the agreement until it’s effectively nothing (except maybe “buy some more of our products”) does generate security value. The real value will come from the hard work: resourcing commitments or other changes to the status quo that move the needle for practitioners. Anything before that is, unfortunately, primarily platitude.

If you’re building in this space, we’d like to hear from you.

Feel free to reach out to jpark@forgepointcap.com and kshih@forgepointcap.com.

This blog is also published on Margin of Safety, Jimmy and Kathryn’s Substack, as they research the practical sides of security + AI so you don’t have to.