Big day. AI found a hole in the plumbing of most of the web, Anthropic got caught shopping a safety bill it wrote for itself, and Walmart decided AI shopping agents are customers, not burglars. Let's go.
01 of 03
AI Found a PHP Bug Older Than Your Truck
On Wednesday the PHP project shipped emergency patches for two critical flaws, CVE-2026-109618 and CVE-2026-109619, which Wiz is calling PHP2Shell. Both score a 9.2 on CVSS. They are memory-safety bugs in how PHP cleans up after a fatal error, especially an out-of-memory error. A request that crashes can leave stale state behind on the worker, and the next request on that same worker can pick it up and run with it.
The bad part is that your app does not have to be sloppy for this to work. Wiz showed unauthenticated remote code execution on WordPress, Drupal, and Joomla, and says even a plain "Hello World" page on an affected version is exposed. Wiz says every PHP version released in the last decade carries at least one of these bugs, going back as far as PHP 5.4. TechCrunch reports the fix to land on is PHP 8.5.12.
Now the AI part. A red-team shop called Calif, working inside OpenAI's "Patch the Planet" program, found the first issue on August 1 using GPT-5.6 Sol. OpenAI dug in, decided it was worse than Calif first thought, and pulled together a group that included Wiz, the PHP Foundation, Cloudflare, Quarkslab, and U.S. government agencies. Wiz, which Google owns, found the problem on its own on August 19 using its Wiz Atlas agent, Gemini 4 Argon, and GPT-5.6 Sol, then reported to PHP on August 24.
The fallout is real. The NSA and the FBI's Cyber Division both told people to patch, and per Fortune, citing Reuters, some government agencies in the Netherlands and Finland pulled websites offline as a precaution. Palo Alto Networks counted 22.6 million potentially vulnerable instances. Cloudflare CTO Dane Knecht says they are already seeing early attacks.
Here is where I land on it. This is the good version of the AI security story, and it is also the scary version, at the same time. The same models that found a bug humans walked past for over a decade will help somebody else write the exploit by lunch. Wiz said it plainly: the skill barrier is gone. If you run anything on PHP, quit reading my blog and go patch. Or call your host and make them do it.
Sources: Wiz Research, TechCrunch, Fortune
02 of 03
Anthropic Wrote Itself a Safety Bill
Politico got hold of a draft bill Anthropic has been quietly shopping to Congress, and TechCrunch laid out how it works. A spokesperson said it reflects the company's "Advanced AI Framework," which Anthropic has been pitching to state and federal lawmakers since June.
The trigger is a "foreseeable, imminent risk" that a model could kill or seriously injure more than 50 people, or cause more than $1 billion in damage. And the risk has to come from one of three places: building or releasing a biological weapon, the model slipping its developer's controls, or a cyberattack.
The mechanics go like this. Independent evaluators register with the Commerce Department. The AI lab can pay them, as long as the money does not depend on what they find. If an evaluator spots one of those imminent risks, they have 72 hours to tell Commerce. Commerce then has to go ask a judge in Washington, D.C. to halt the release. No federal agency, Commerce included, gets to stop a model on its own say-so.
Critics did not hold back. Heidy Khlaaf, chief AI scientist at the AI Now Institute, called it "autocratic" and said it lets the companies set their own harm bars while ignoring the harms already happening, according to Cybernews. Mistral CEO Arthur Mensch took the opposite swing, saying American labs use doomsday talk to win market share.
What I want to know is who this bill actually protects. Credit where due, it goes further than most labs, and Jensen Huang is out there saying we need no new AI laws at all. But read it again. The company gets to pay its own graders, the government cannot act without a court, and a bioweapon risk to 49 people does not count. That is not a leash. That is a very nice dog bed with Anthropic's name stitched on it.
03 of 03
Walmart Rolls Out the Welcome Mat for AI Agents
While Amazon is busy blocking bots, Walmart went the other way. CEO John Furner announced it Wednesday at the Masters of Scale Summit in San Francisco, per TechCrunch, and there are three pieces.
First, Agent Gateway, an MCP server that lets approved AI agents search Walmart's inventory and buy for you. It is invite-only beta for now. Second, Agent Profiles, coming within the next month. Each agent gets its own slot in your account so its purchases show up separately. You can make it ask permission every time, or let it spend up to a limit you set. Walmart threw out $50, $100, and $200 as examples, per PPC Land. You can delete a profile anytime, and you do not need Walmart+. Third, a set of rules: agents cannot touch your card details, cannot go past what you asked for, and get blocked if they break the terms. If an agent screws up, normal Walmart returns still apply. It all covers Sam's Club too.
The contrast is the story. Last month Amazon blocked Meta's Muse agent from buying on Amazon.com, saying it never agreed to Muse shopping its store and that Muse did not identify itself. Walmart already lists Muse among its integrations, alongside ChatGPT and Gemini.
I think Walmart read this one right. Your customers are going to send robots shopping whether you like it or not, so you might as well set the house rules and take the sale. The spending cap plus a separate profile is exactly the "authorized user on the credit card" model normal people already understand. Amazon can keep its fence up. I suspect it will not stay up for long.