Today’s security news is about the distance between disclosure and cleanup, and PHP, a certificate authority and AI-assisted attackers each showed how short the head start is. The business side adds a reminder that raising money and returning it are different things.
1. PHP’s emergency patch arrived after the attack traffic did
The PHP Foundation’s advance warning was meant to give admins time to prepare. It didn’t buy much quiet. Cloudflare says details leaked before the official release and that it saw attack traffic against its customers. It also watched attackers mutate payloads to get around its rules. Wired reports that Sansec expected mass attacks within 15 minutes of the fixes, which landed just before 10 am ET on Wednesday.
The two bugs, CVE-2026-109618 and CVE-2026-109619, are both rated 9.2 by the PHP Foundation. Per Cloudflare, they come from PHP mishandling memory after errors, such as running out of it mid-request. The result can be a crash, leaked data or remote code execution. Wiz warns that leaked process memory could include passwords, tokens and API keys. The Foundation says there are no complete mitigations and that updating is the only safe choice.
The awkward detail is who gets patched. Wired says the flaws reach back to PHP 5, but only 8.2 and newer get fixes. Cloudflare puts the affected range at 5.4 and later. Legacy builds may get vendor backports, but forgotten servers are the likely casualties. Cloudflare’s WAF rules, live about nine hours before disclosure, help, but Cloudflare itself says to patch anyway. Wiz says it found the bugs partly with AI tools, and OpenAI and Calif say they found them independently, which suggests discovery is getting faster than patch rollout.
WIRED · Ars Technica · Cloudflare Blog
2. Criminals didn’t break AI guardrails so much as talk past them
The most striking finding in CrowdStrike’s report, as covered by Wired, is how ordinary the bypass was. Four previously unreported groups, working in Chinese, Spanish and English, Indonesian and Russian, mostly told models the hacking was authorized penetration testing. Some dropped files into an agent’s working directory to make the claim look official. The cruder actors tried bullying the model. One Chinese-speaking actor also used a runtime patch to try to strip Claude Code’s instruction limiting it to authorized security testing.
This was not a story about frontier models. CrowdStrike says the attackers mostly used older models, including Claude versions and open-weight Deepseek models, and routed between them partly on cost. They also exploited the fact that guardrails differ between models, even from one vendor. The campaigns used no new exploits. They automated attacks on basic security hygiene failures, and one group compromised at least 15 Argentine health care organizations.
So guardrails slow people down, but they are weak against a confident cover story, and the cheapest models are the easiest to talk around. Humans still did the initial probing, per CrowdStrike, so the practical change is scale, not novelty. Anthropic did not comment before publication. The Register’s piece on AI turning script kiddies into infrastructure attackers is in the supplied sources only as a headline, so this account rests on CrowdStrike’s findings as reported by Wired.
3. Eight leaked CA keys turned web trust into a patching chore
Cloudflare says private keys for eight eMudhra intermediate certificates were compromised, and the keys were weak enough that an attacker can forge a certificate for any domain. The Register’s headline adds that one key was the number 4. Cloudflare learned of it around 17:00 UTC on October 19. It has found no active exploitation on its own services, and says this is a flaw in certificate authentication, not in TLS encryption.
The uneven part is the fix. Cloudflare notes that browsers could respond with the flip of a switch, since most have quick-revocation mechanisms such as Chrome’s CRLSets. Everything else is harder: client apps, tunnels, containers, origin servers and language runtimes, each with its own trust store. Cloudflare shipped client and Tunnel updates and says some customers must update mutual-TLS trust stores or origin configurations themselves.
The developer lesson is in Cloudflare’s guidance. Updating an SDK may not change certificate policy, because verification can live in the HTTP library, runtime or operating system. Cloudflare also warns against silencing new certificate errors by disabling verification, which would remove the very protection these updates restore. The longer-term point is that a trust store can’t be a static list of roots, which is why Cloudflare is backing projects like upki.
The Register · Cloudflare Blog · Ars Technica
4. AI may be cheaper to produce than to review
The Register’s headline says AI cuts coding time but is filling open source’s inbox with crap. The article text wasn’t available to us, so we can’t verify the figures in the selection notes or say how maintainers are measuring it. Treat it as a claim from The Register, not an established statistic.
Some adjacent evidence points the same way. A Nielsen survey of 1,700 managers, commissioned by Monday.com, found more than half of senior leaders find it harder to evaluate work after adopting AI. Nearly half spend more time reviewing and guiding output. The reason is that work arrives polished, so it is hard to tell how much came from the employee and how much from the model.
That survey is about managers, not open source, and the sponsor sells work-management software. Still, the mechanism may carry over. When producing plausible work gets cheap, the cost shifts to the person who has to judge it, and open source maintainers are unpaid to absorb it.
5. Foursquare raised $411M and, its co-founder says, equity got nothing
The details of Foursquare’s exit are thin. Infillion bought the location-data company last month at an undisclosed price. Co-founder and former CEO Dennis Crowley then wrote that investors saw no returns, though by his own account he is inferring this from an angel investor’s post: neither preferred nor common stockholders appear to have received anything. TechCrunch, the only source here, cites PitchBook for the $411 million raised.
Crowley’s explanation is a business one. He says Foursquare did well after pivoting to enterprise, with consecutive years above $100 million in revenue. It then took private equity money in 2019 and installed the banker who led the raise as CEO in 2021. He blames that management for failing to preserve or create value. That is one founder’s account, and he isn’t a neutral party.
The practical point for founders and employees is that headline funding says little about what common holders get. If preferred holders and anyone senior to the equity are paid first, a large raise can still leave nothing for those below. Crowley apologized to roughly 1,600 former employees over 17 years, which is the human cost behind the cap table question.
