Megan Skiendel: David, I've been sitting with this all week and I think the cleanest way I can put it is — what if 'audited' never actually meant what we thought it meant?
David Sterling: Unpack that. What are you pointing at?
Megan Skiendel: PHP. Seventy percent of the web runs on it. Human reviewers had eyes on that codebase for over ten years — version control, security passes, the whole apparatus. And a memory disclosure bug plus an unauthenticated remote code execution flaw sat inside every version going back to 5.4. Untouched. Because humans spot-check. We do not read exhaustively.
David Sterling: And a machine did.
Megan Skiendel: GPT-5.6 Sol — running through Project Daybreak, which is OpenAI's elite security research community, all under the Patch the Planet initiative — worked through that code the way no human review team ever could. Not faster, necessarily. Just without stopping. Without skipping. Wiz got there independently too, reported both vulnerabilities in parallel. Which tells you the AI-assisted methodology isn't one team getting lucky.
David Sterling: It's reproducible. That's actually the more alarming part — if two teams can find it, so can a third team with worse intentions.
Megan Skiendel: That is exactly the tension this whole episode is about. And it starts with one question: does finding something faster actually help you, if the seventy percent of the web running PHP patches at the speed it patches?
David Sterling: Well. That's the number I want to get to.
Megan Skiendel: Patching speed is where the coordinated disclosure protocol completely falls apart under pressure. October 21st — PHP ships the patches. That's OpenAI, Wiz, Cloudflare, all in the room, doing this correctly. Two days. Forty-eight hours later, an unnamed cybersecurity company just... publishes. Full flaw details. Before operators have caught up.
David Sterling: Hold on. Forty-eight hours?
Megan Skiendel: October 23rd. That's the number. And the coordinated disclosure protocol — which exists specifically to prevent this — assumed a world where discovery is slow and the patch window is measured in weeks, not days. AI just compressed the timeline past the point where those governance structures were designed to hold.
David Sterling: I mean — whoever made that call, do we even know if it was malicious? Or was it someone calculating that the vulnerability was already being exploited and disclosure accelerates defense faster than it arms attackers?
Megan Skiendel: Honestly, we don't know. It could be competitive pressure — claim the discovery credit before someone else does. It could be genuine disagreement with the timeline. But either way, the cascade from that one decision: sites going fully offline rather than wait for a patch window, the FBI's cyber division issuing emergency warnings, PHP maintainers scrambling. And it wasn't just PHP by then.
David Sterling: Node.js, Plex, LiteLLM, Railway, eMudhra — all pulled into it.
Megan Skiendel: Which tells you the blast radius of that premature disclosure isn't one flaw — it's every maintainer watching PHP go sideways and wondering if they're next. The FBI warning lands and suddenly the ops teams who were planning a controlled patch window are taking systems offline at 6 a.m. rather than risk being the cautionary tale.
David Sterling: And Cloudflare — this is the part that actually surprises me — they contributed to the coordination. They saw the gap from the edge. Every vulnerable installation proxying through their network, and they had zero authority to force a patch. They could see it. Couldn't touch it.
Megan Skiendel: That's the structural problem with coordinated disclosure in an AI-speed world — the people with the clearest view of the exposure surface aren't the ones with the keys. Chris Krebs said it plainly: constantly scan, patch, or take legacy systems offline. But that assumes the organization has a window to make that call before someone else makes it for them.
David Sterling: Which means the actual question isn't whether AI found the flaw. It's whether this is a one-time feat or a permanent new operating condition.
Megan Skiendel: Industrial. That's the word. Microsoft FORGE — May through September 2026 — 140 Windows CVEs. One hundred fifty-five validated reports across twenty-three open-source projects, the Linux kernel included. That's not a research project. That's a production line.
David Sterling: 140 CVEs in five months.
Megan Skiendel: And ASF — the Apache Software Foundation — ran Project Glasswing in August 2026. Anthropic's Claude Mythos 5, scanning 230 repositories. Three days. A human team doing that exhaustively would take... I mean, honestly, months? Maybe never finishes? The pace is categorically different.
David Sterling: But if defenders can scan 230 repos in three days, what's the attacker's equivalent timeline?
Megan Skiendel: That's the dual-use problem stated plainly. The same capability. Anthropic's own opt-in vulnerability service — nearly 5,000 unvalidated reports sent directly to open-source maintainers. And Anthropic's own citation is that exploits can now be developed in minutes. So the discovery is fast and the weaponization is fast. Both sides of that equation accelerated simultaneously.
David Sterling: VulnCheck says AI-assisted discoveries doubled year-on-year. Do we actually know the baseline? Because doubled from what?
Megan Skiendel: Treat it directional, not precise. The methodology is thin. But Okta publishing internal findings on frontier AI for vulnerability hunting — that's not a number, that's a signal. When Okta documents it, every mid-market security team reads that and starts asking their vendor about access. The practice spreads faster than the data does.
David Sterling: And the part that makes all of this genuinely unsettling — we'll get to it — is that defenders have to patch every single instance. An attacker needs exactly one working exploit. That asymmetry is where this whole acceleration story actually lands.
Megan Skiendel: And that asymmetry — every instance versus one exploit — it doesn't dissolve just because discovery got faster. It gets worse. Because now the window between disclosure and weaponization is minutes, not weeks. Anthropic said that explicitly. Minutes.
David Sterling: So Project Daybreak and Project Glasswing give defenders a head start. But the moment coordinated disclosure breaks — October 23rd — that head start is just gone.
Megan Skiendel: Instantly gone. And then you get the actual human version of this. Picture a Friday afternoon — an e-commerce ops team sees the patch notification come in. They want to deploy. Their CTO is unreachable. Their PHP stack hasn't been touched in eighteen months. They are not exposed because they didn't know. They're exposed because patching is — I mean, it's not a technical act. It's an organizational one. Someone has to authorize downtime.
David Sterling: That's exactly it. The technical fix exists. The organizational chain doesn't.
Megan Skiendel: Which is why Chris Krebs telling organizations to constantly scan and patch or take legacy systems offline — that's not advice. That's an admission.
David Sterling: An admission of what, exactly?
Megan Skiendel: That the system is already priced in organizational failure. The former CISA director is not saying 'here's how you succeed.' He's saying 'here's the minimum floor to not be the cautionary tale.' And forty percent of PHP installs are still unpatched — after all of it. The emergency warnings, the coordinated response, the FBI. Forty percent.
David Sterling: Well. That's the number that doesn't move. Discovery is solved. Remediation isn't even close.
Megan Skiendel: And that's — honestly, that's the detail that sticks. Not the AI capability. The CTO who's in a meeting while the exploit window is already open.
David Sterling: And the machine read all of it. GPT-5.6 Sol read every line, found the flaw, told us. And then we had forty-eight hours before someone made a room decision and the window closed.
Megan Skiendel: The machine did its job. Completely. And then it handed the problem to the part of the system that hasn't changed — the CTO in a meeting, the eighteen-month-old PHP stack nobody wants to touch, the seventy percent of the web that patches on its own schedule regardless of what any AI found or when. That dependency didn't move.
David Sterling: Cloudflare sitting at the edge. Watching every vulnerable install route through. Couldn't patch a single one.
Megan Skiendel: That's the image that doesn't leave me. The next crisis — honestly, I don't think it gets solved by a better model than Sol. It gets solved by whoever can actually compel legacy systems offline or compliant before the weaponization window opens. That's infrastructure control. That's not a research problem.
David Sterling: We started with whether 'audited' meant what we thought it meant. Turns out the audit was never the weak point. Appreciated the time on this one.
