AI News: October 22, 2026

1. PHP Patches Critical Flaws in Every Version Since PHP 5, Found With GPT-5.6 Sol and Gemini 4 Argon

PHP Foundation. The PHP Foundation shipped patches for two critical flaws that reportedly affect every PHP release since version 5 and can give attackers access to a server’s file system or backing database; the NSA issued a warning and recommends PHP 8.5.12. Security firm Calif found one flaw with GPT-5.6 Sol under OpenAI’s Patch the Planet program, and Google’s Wiz team found several using its Wiz Atlas agent, Gemini 4 Argon, and GPT-5.6 Sol. Disclosure was coordinated and no in-the-wild exploits are known, but WordPress and other PHP hosts still need to patch their servers. Source

2. Recursive Releases Flint, a 40B Open-Weight Coding Model Built Largely by Its AI Research System

Recursive. Richard Socher’s Recursive released Flint, a 40B-parameter mixture-of-experts coding model with about 2.5B active parameters, under Apache 2.0 on Hugging Face. The company says it leads models under 50B parameters on SWE-bench Verified and runs at about 131 tokens per second on an M4 Pro MacBook Pro. Recursive says its automated “Eureka Machine” handled nearly every development step, including a sparser architecture and low-bit training that cut model size by 71%. Source

3. Inception’s Mercury 3 Diffusion Model Claims 4-8x Speed Over GPT-6 Luna and Haiku 5.5

Inception. Inception’s Mercury 3 diffusion language model generates about 1,068 tokens per second on Nvidia GPUs, versus roughly 240 for Claude Haiku 5.5 and 130 for GPT-6 Luna, and the company says it beats both on 8 of 12 benchmarks, including 51.1 on SWE-Bench Atlas. It has a 1M-token context window and costs $0.20/$0.70 per million input/output tokens. No independent evaluations have been published yet. Source

4. Flapping Airplanes Raises $1 Billion at a Reported $6 Billion Valuation

Flapping Airplanes. The year-old research lab raised a $1 billion Series A co-led by Index Ventures and Lux Capital, with Bloomberg reporting a $6 billion valuation, about four times its January seed valuation. The lab, led by ThunderKittens author Ben Spector, aims to build models that need far less training data and has spent its first year on a “megakernel” framework that runs a whole program as a single GPU kernel. It has no commercial product yet. Source

5. Blacksmith Raises $66 Million at $1 Billion as AI Code Strains CI Pipelines

Blacksmith. CI platform Blacksmith raised a $66 million Series C led by Sequoia at a $1 billion valuation, two months after a Series B at $550 million. The company says ARR grew 17x and it has run more than 90 million CI jobs for customers including Mercury and Granola. The round reflects how the volume of AI-generated code is shifting bottlenecks to build and test infrastructure. Source

6. Walmart Opens an MCP Agent Gateway for AI Shopping Agents as Amazon Blocks Them

Walmart. Walmart launched Agent Gateway, an MCP server that lets approved AI agents search inventory and buy from Walmart and Sam’s Club for a customer without seeing payment details or full purchase history. Agent Profiles, rolling out next month, give each agent its own entry in the account with per-transaction approval or a preset spending limit, and agents that exceed the shopper’s intent will be blocked. The move contrasts with Amazon, which has begun blocking shopping agents from its site. Source

7. Red Hat and Nvidia Move Agent Security Controls Out of the Agent and Into the DPU

Red Hat. Red Hat expanded Red Hat AI Factory with Nvidia to run Nvidia’s Sentry framework on BlueField DPUs, which watches agent processes, files, and network traffic from a separate hardware domain and can quarantine a compromised agent without host-side software. A joint reference design pairs model-as-a-service access controls and an AI gateway with vLLM, llm-d, Dynamo, and NIM, and adds Nemotron 3.5 Super VL support in vLLM. Native OpenShell integration is still in progress; the designs are available through Cisco, Dell, and Lenovo. Source

8. China Demands the US Lift Its DeepSeek Ban and Threatens Countermeasures

China. China’s Commerce Ministry demanded the US lift the DHS order removing DeepSeek from federal civilian systems and promised unspecified “resolute countermeasures,” while the Foreign Ministry accused Washington of stretching its national security rationale. Beijing itself blocks ChatGPT, Claude, and Gemini nationwide. The response signals possible retaliation that could affect US AI and chip companies operating in China. Source

9. DHS’s Revised AI Directive Lets Its Chief AI Officer Waive Any Requirement

DHS. The Department of Homeland Security’s revised AI directive (139-08, Revision 01) lets the chief AI officer waive any of its requirements, including prohibited use cases, which civil liberties groups called a loophole. The directive also requires DHS to release custom AI code as open source in a public repository and share models across government, while removing language on equity and environmental sustainability. It keeps the ban on large-scale surveillance and emotion or future-behavior prediction largely intact. Source

10. Judge Rejects Bid to Let EPA Dismiss NAACP Suit Over SpaceXAI’s Colossus Turbines

SpaceXAI. A federal judge rejected the Trump administration’s argument that the EPA can unilaterally dismiss citizen Clean Air Act suits, keeping alive the NAACP’s lawsuit over unpermitted gas turbines powering SpaceXAI’s Colossus data center in Memphis. The suit seeks best-available pollution controls and penalties of up to $124,426 per day per violation. The ruling preserves a legal avenue for communities challenging fast-tracked AI data center power. Source

11. OpenEvidence Reaches a $15 Billion Valuation With Nearly Two-Thirds of US Doctors Using It

OpenEvidence. OpenEvidence, an ad-supported medical search chatbot, says US physicians sent it more than 42 million prompts last month, and nearly two-thirds of US doctors now use it to consult journals and guidelines for diagnosis and treatment. The company is valued at $15 billion. Its reach raises questions about accuracy and advertising influence in a tool that sits inside clinical decision-making. Source

12. Hassabis Shortens His AGI Estimate to Two to Three Years

Demis Hassabis. In a Wall Street Journal interview, DeepMind founder Demis Hassabis said AGI, defined as a system matching the full cognitive range of the human brain, could arrive within two to three years, down from the three to five years he estimated earlier this year. “My estimate keeps moving closer to the present,” he said. Hassabis stepped back from Google DeepMind and now focuses on Isomorphic Labs. Source

13. Deel Launches Pangaea, Small Specialist Models for HR and Payroll Questions

Deel. Deel released Pangaea K1, a 9B-parameter employment-rules model trained on 43,500 pages of specialist material, and P1, a Nemotron-based payroll model further trained by AI21 Labs. On Deel’s own 601-question EmBench Global Work, K1 scored 70 against 74 for Claude Opus 5.5 and 75 for DeepSeek V4.1 Flash, with a 7.7-second median response time and roughly 1/49th the cost per answer. The models run inside Deel AI only; no weights or public API were announced. Source

14. CTGT Open-Sources Concordance to Find Where Agent Runs Diverge

CTGT. Interpretability startup CTGT released Concordance under Apache 2.0, a protocol and reference implementation that captures exact model requests and responses and finds the earliest point where repeated runs of an AI application diverge. In CTGT’s test, a compliance app extracted 31 to 33 obligations from the same banking manual across four runs even though the model answered all 74 identical requests consistently; the variance came from upload IDs, an unsorted routing list, and a growing topic registry. After fixing those, five runs each produced the same 34 obligations. Source

15. ScholarEvolve Turns Research Papers Into Agent Harness Upgrades Without Retraining

UC Santa Barbara and Microsoft. Researchers introduced ScholarEvolve, which reads academic papers and converts their ideas into code changes to an agent’s harness while leaving model weights untouched. It raised Qwen3.5-27B from 69.0% to 81.4% on AppWorld Normal and from 49.6% to 63.6% on AppWorld Challenge, 9 points ahead of Stanford’s Meta-Harness, and lifted GPT-5.4-mini from 72.7% to 81.9% on Tau-Bench Telecom. Source

16. Survey: Over Half of Managers Find It Harder to Evaluate Work Since Adopting AI

Monday.com. A Nielsen survey of 1,700 directors and senior managers in seven countries, commissioned by Monday.com, found that more than half find it harder to evaluate employees’ work since adopting AI and nearly half spend more time reviewing output. 58% expect to oversee more AI agents than human employees within a year, while 46% worry about losing control over agent actions. Only 29% describe their AI infrastructure as fully optimized and centralized. Source

17. Scaled Cognition Releases a $9-per-Million-Character TTS Model for Customer Service Agents

Scaled Cognition. Scaled Cognition launched Voice, a text-to-speech model tuned to read account numbers, addresses, and amounts correctly, for example rendering “Pay 1/2 by 1/2” as “pay half by January second.” It costs $9 per million characters hosted or $1 self-hosted (excluding infrastructure), versus $35 to $100 the company cites for rivals, with about 70ms latency and support for 25+ languages. Accuracy claims rest on blind listening tests with no published sample sizes. Source