Disclosure is outrunning patching, and Anthropic's answer is to prove code correct instead of testing it harder.

Today's headline

Disclosure is outrunning patching. PHP has flagged a third remote code path with no fix scheduled. Pillow shipped a batch of critical image fixes, and Next.js has an unauthenticated remote code bug with no workaround. Anthropic's answer is to prove code correct, not test it harder.

On the clock

Today, OpenAI retires a long list of older models, including gpt-4 and o1. Google's older Deep Research agent shuts down the same day. MiniMax's Flash model is expected to go open weights, and SGLang already supports it.

On the thirty-first, OpenAI Evals goes read-only and DeepSeek's free search promotion ends. Mistral retires one of its GLM models the same day.

Do this week: repoint the retiring OpenAI models and export your saved evals.

Security

The lead is PHP. Two days after its emergency fix, PHP warned of a third way to run code. It is narrow. It needs a specific session setting changed at runtime, with upload progress on. There is no emergency release, no severity score yet, and scanners will be silent. No attacks on it are reported. The earlier two flaws are another story. Amazon logged an attempt four days before disclosure, and Cloudflare saw attempts reach the millions within a day. The fix for the new one is a single line: turn off session upload progress, and keep the earlier emergency patch in place.

Next.js has a critical, unauthenticated remote code bug. It hits App Router apps with a cached function that returns typed arrays. There is no workaround, and no exploitation is reported. The maintainers say to upgrade even if you do not have such a function.

Pillow version thirteen fixes a batch of critical-rated flaws. One lets a crafted filename run commands through Ghostscript. Another was found by Anthropic using Claude. No attacks are known. The same release also fixes a TIFF overflow that has a public proof of concept. Upgrade, and keep Ghostscript off hosts that process uploads. Python three ten installs need to move to three eleven first.

And the x402 payment libraries let a malicious server ask an agent to sign for its whole token balance. It was fixed in August, but the project filed no advisory. Your audit tools will say nothing.

Do this week: upgrade Pillow, set the PHP ini value, and patch Next.js.

Agents

Anthropic and Theorem used Claude to formally verify one hundred twenty-three open-source libraries. That is about one hundred eighty million lines of proof, checked automatically. Anthropic says the proofs hold only for one processor type, single-threaded code and smaller inputs. Single proofs can run to hundreds of millions of tokens. During the work, Claude found hundreds of bugs that testing had missed. It is Anthropic's own result. Anthropic expects this to become as routine as testing, and Theorem's roadmap names glibc, OpenSSL and curl.

US banks are locking out agents that log in with customer credentials. Morgan Stanley blocks them outright, and JPMorgan, Bank of America, Wells Fargo and Citi are restricting them. Most banks want read-only access, and most cannot tell an agent from its owner. On the same day, 1Password launched a way for agents to sign in with the user approving each login, and the agent never sees the password.

Google previewed Agent Compute Engine, which gives each agent its own isolated virtual machine. Google says one server can hold ten times as many agents.

Straiker reports a crime group used Claude, Grok and Moonshot's agent against hundreds of Latin American institutions. Anthropic has blocked the access.

Models and pricing

OpenAI made the Decisions API generally available. It returns a choice and a confidence score, and OpenAI says calls now take about one hundred to one hundred eighty milliseconds. Speak, an early customer, reports calls falling from one point four seconds to a tenth of a second. It is a reconfigured Luna, not a purpose-trained model, so test it against your own workload before you swap it in for an LLM call.

OpenAI also added twelve-hour prompt caching. Writes cost two and a half times the input price, so it pays off only when a prefix is reused across hours. Idle sandbox time is no longer billed.

DeepSeek's latest Flash update beat Claude Haiku in one hands-on test at a lower price. DeepSeek charges about fifteen cents in and sixty cents out per million tokens off-peak, and double at peak. Both benchmark scores are vendor-reported, and Tom's Hardware finds it hallucinates more than proprietary rivals.

OpenAI donated Open Responses, and Anthropic donated Agent Skills, to the Agentic AI Foundation.

Business and policy

OpenAI is reportedly raising up to thirty billion dollars at a valuation of one point two five trillion. Terms are not settled. DeepSeek reportedly raised about fifteen billion, which it has not confirmed.

A Harvard study of seven hundred eighteen firms found agents raise lines of code by about thirty percent, but resolved issues rise only twelve percent. Code review absorbs most of the gain. Plain assistants did not raise resolved issues at all.

A New Mexico court ordered Meta to pay twenty-five billion dollars over privacy violations tied to Cambridge Analytica. And Lava says it found thirty-two export-barred Nvidia B300 chips inside China Telecom's network. Nvidia called diversion a drop in the bucket and did not deny the finding.

Do this week

Upgrade Pillow to version thirteen and rebuild your images. Set the PHP session upload ini value to zero on every host. Patch Next.js to the fixed release. Confirm your x402 agents have spending caps. Export your OpenAI Evals before the thirty-first.

Before you go

Check your Python images first today. Pillow sits under almost everything that touches a picture, and that is the fix with the widest reach.