OpenAI's security pivot looks familiar, and that's actually the point

The Promptfoo acquisition is the most concrete move here. Promptfoo built tools that help enterprises test their AI systems for vulnerabilities during development, not after deployment. Think of it as security testing shifted left, to use the DevOps jargon the kids love. The acquisition suggests OpenAI wants to bake security testing into the development pipeline rather than bolting it on afterward. Whether they'll make Promptfoo's tools available to customers or keep them internal remains unclear.

The October 2025 disruption report is worth reading if you have the patience. It documents specific cases where OpenAI detected and shut down malicious actors using their models, including influence operations, fraud attempts, and what they call "real-world harms." The report is light on technical specifics (understandably, you don't want to give bad actors a roadmap), but it does show they're actually monitoring for abuse rather than just hoping it doesn't happen.

Why now, and why all at once?

Call me old-fashioned, but I think the timing here is about liability as much as ethics.

AI agents are getting more capable. OpenAI's models can now browse the web, execute code, and take actions on behalf of users. That's useful! It's also terrifying from a security perspective. An AI agent that can send emails and make purchases on your behalf is also an AI agent that can be tricked into sending emails and making purchases for someone else.

The cyber resilience post from OpenAI acknowledges this directly. They write about assessing risk as models become more powerful, limiting misuse, and working with the security community. It's corporate speak, sure, but the underlying message is clear: they know their models could be weaponized, and they're trying to get ahead of the inevitable regulatory and legal scrutiny.

I've seen this movie before. Microsoft's "Trustworthy Computing" memo in 2002 came after years of security disasters, and it took them another decade to actually become reasonably secure. The difference is that AI systems can cause harm much faster and at much larger scale than buggy operating systems ever could. OpenAI doesn't have a decade to figure this out.

The EVMbench thing is weird, but interesting

The EVMbench announcement feels like it belongs in a different conversation, but it's actually relevant. OpenAI partnered with Paradigm (a crypto venture firm) to create a benchmark that tests whether AI agents can detect, patch, and exploit vulnerabilities in Ethereum smart contracts.

Why does an AI company care about smart contract security? Because smart contracts are basically autonomous agents that handle money, and they're a perfect test case for AI-assisted security research. If an AI can find bugs in smart contracts faster than humans can, that's potentially valuable. If an AI can exploit smart contracts, that's potentially catastrophic.

The benchmark is designed to measure both capabilities. OpenAI says they want to understand how AI can help with defensive security, but they're also clearly mapping offensive capabilities. You can't defend against AI-assisted attacks if you don't understand what those attacks look like.

Does any of this actually work?

Here's where I have to be honest: we don't know yet.

Bug bounties are only as good as the researchers who participate and the company's willingness to actually fix what gets reported. OpenAI's track record here is, well, mixed. They've had public disputes with researchers over disclosure timelines and bounty amounts. The Safety Bug Bounty might be different, but I'll believe it when I see sustained participation from top-tier security researchers.

The Promptfoo acquisition could be transformative or it could be talent acquisition that gets buried in a larger organization. I've watched plenty of promising security startups get acquired and then basically disappear. The coordinated disclosure policy is a good sign that OpenAI wants to be a responsible participant in the security ecosystem, but policies are easy to write and harder to follow.

And the disruption reports are helpful for transparency, but they're also self-reported. We're trusting OpenAI to tell us how well OpenAI is doing at catching bad actors. That's not nothing, but it's not independent verification either.

So what should we actually take from this?

Look, I'm not going to tell you OpenAI has solved AI security. They haven't, and anyone who claims otherwise is selling something. But I will say this: the company appears to be taking security seriously in a way that feels substantive rather than performative.

The combination of internal investment (Promptfoo), external engagement (bug bounties, disclosure policies), public transparency (disruption reports), and capability measurement (EVMbench) suggests a coordinated strategy rather than a PR response to some crisis we don't know about. Maybe there is a crisis we don't know about! But the approach looks right regardless.

The question I keep coming back to is whether other AI companies will follow. Anthropic has been relatively quiet on security specifics. Google has DeepMind's safety work but less visible security infrastructure. Meta is, in a way, open-sourcing the problem by releasing model weights and letting the community figure it out.

If OpenAI's security push becomes an industry standard, that's good for everyone. If it remains a competitive differentiator that other companies ignore, we're going to have problems. Big problems, and probably sooner than anyone wants to admit.

I've been covering tech long enough to know that security only becomes a priority after something goes badly wrong. The smart money says OpenAI is trying to get ahead of that moment. Whether they succeed, well, that's a question we won't be able to answer for years. But at least they're asking it.

If you want to argue about any of this, my email's on the about page.

Mark Kowalski · 1 hour ago · 5 min