AI News · October 22, 2026 · 7:17

AI-assisted exploits hit PHP & AI safety enters the mainstream - AI News (Oct 22, 2026)

AI helps hack PHP, DeepMind staff quit over military work, OpenAI maps the brain in 32 minutes, and Meta's Muse explodes. Your daily AI briefing.

English Français Español
AI-assisted exploits hit PHP & AI safety enters the mainstream - AI News (Oct 22, 2026)
0:007:17

15 free episodes a day, 5 editions, 3 languages, about $4 a day to produce. A coffee keeps it going ☕

Support the show →

Our Sponsors

SurveyMonkey, Using AI to surface insights faster and reduce manual analysis time Learn more → Invest Like the Pros with StockMVP Learn more → KrispCall: Agentic Cloud Telephony Learn more →

Today's AI News Topics

  1. AI-assisted exploits hit PHP

    — Wiz Research found critical memory-safety flaws in the PHP engine affecting versions back to PHP 5.4, with frontier AI tools helping turn old bugs into remote code execution exploits. PHP shipped patches on October 21, 2026.

  2. AI safety enters the mainstream

    — Rufus Knuppel argues Jacob Coxon's resignation from Anthropic pushed AI safety, effective altruism and rationalist concerns from a Bay Area subculture into mainstream American politics.

  3. DeepMind staff protest military AI

    — Current and former Google DeepMind employees are resigning or speaking out over military AI work for U.S. and Israeli defense projects, citing ethics and battlefield escalation risks, according to the BBC.

  4. OpenAI's new monitorability metrics

    — OpenAI Alignment Research proposes pair-matched comparisons and post-control risk metrics to complement g-mean² when measuring AI monitorability and deployment risk.

  5. Pichai on frontier AI and compute

    — Sundar Pichai says Gemini 4 Argon reflects Google's frontier AI commitment, discusses recursive self-improvement, world models, robotics and aggressive data center and chip investment amid compute shortages.

  6. Google and Meta launch TorchTPU

    — TorchTPU, an open-source collaboration between Google and Meta, gives PyTorch a native TPU backend with minimal code changes, torch.compile support and compatibility with FSDP, vLLM and Hugging Face.

  7. AI agents map brain CRISPR data

    — OpenAI, PerturbAI and Scripps Research used thousands of GPT-Rosalind agents to analyze a whole-brain mouse CRISPR dataset in 32 minutes, surfacing hypotheses on Tsc1, Tsc2, cholesterol and neurodegeneration.

  8. OpenAI acquires Orchestra Bio

    — OpenAI acquired Orchestra Bio, an AI platform for coordinating drug development, to combine with its Rosalind life sciences models and build agents that speed the path from discovery to treatment.

  9. Meta's Muse app growth surge

    — Meta's AI companion app Muse hit 10 million U.S. downloads in 40 days, per Sensor Tower, with 2.5 million daily active users and stronger retention than ChatGPT, Claude and Grok.

  10. ChatGPT adds Locked Chats

    — ChatGPT's new Locked Chats feature protects individual conversations with a PIN or Face ID, separating them from history and search, though it does not add end-to-end encryption.

  11. DeepSeek and llm-d serving stacks

    — DeepSeek open-sourced DeepInfer, its distributed inference framework for DeepSeek V4.1 Flash with long-lived KV cache, while llm-d reached version 1.0 as a Kubernetes-native LLM inference stack.

  12. Durable infrastructure for AI agents

    — Vercel released open-source durable-harness so coding agents like Claude Code and Codex survive crashes, Ramp detailed its agent-built Laminar platform, and Databricks introduced MemCurrent persistent agent memory.

  13. Controlling AI coding tool costs

    — Databricks argues rising AI coding tool costs can be controlled through cheaper and open models, routing, caching, token reduction and spend governance via an AI gateway.
  14. 14

    Mistral Voxtral 3 voice agents

    — Mistral launched Voxtral 3, a multilingual real-time voice agent platform combining speech-to-text, text-to-speech and an agent SDK with EU-hosted and self-deployment options.
  15. 15

    Small on-device model Spark-X2.6

    — XHToken released Spark-X2.6-4B, an Apache 2.0 on-device language model with configurable thinking levels and improved coding, agentic and long-context performance.
  16. 16

    Limits of AI hardware engineering

    — An essay argues AI may soon reach superhuman hardware design, but manufacturing, testing and real-world iteration remain the hard bottlenecks for hardware engineering.

Sources & AI News References

Full Episode Transcript: AI-assisted exploits hit PHP & AI safety enters the mainstream

A cluster of bugs hiding inside the software that runs a huge slice of the web, some of them for over a decade, just became practical attacks. And researchers say AI is what made the difference. Welcome to The Automated Daily, AI News edition. The podcast created by generative AI. I'm TrendTeller, and today is October 22nd, 2026. Let's get into it.

AI-assisted exploits hit PHP

We start with that security story. Wiz Research says it found several critical memory-safety flaws in PHP's core engine, affecting nearly every version released in the past decade. The problem isn't WordPress, Drupal or Joomla themselves, but the language underneath them, so even simple PHP apps could be exposed. Wiz showed attacks that could pull passwords and API keys out of memory and, in some cases, run code remotely. PHP shipped patches yesterday. The bigger takeaway is the role of frontier AI: Wiz says these were long-known, hard-to-exploit bugs that AI tools helped turn into working exploits far faster. If AI is lowering the bar for attackers, defenders will need to patch and respond a lot quicker.

AI safety enters the mainstream

On to AI safety and the culture around it. A feature by Rufus Knuppel argues that Jacob Coxon's resignation from Anthropic was the moment AI safety left its Bay Area niche and entered mainstream American debate. The piece paints Berkeley's Constellation hub and the effective altruism and rationalist communities as intense and ideologically driven, and highlights an awkward split: some want to slow AI down, while others are building the very systems they fear. Its point is that AI is now treated as a civilizational contest, not just an engineering project.

DeepMind staff protest military AI

That tension is showing up inside the labs, too. The BBC reports that current and former Google DeepMind employees are resigning or speaking out over the company's growing military work, especially U.S. and Israeli defense-related projects. One research engineer said Google no longer matched her ethical values. Staff worry AI could make battlefield decisions faster than humans can check them, and some say their contracts discourage public criticism. Google points to its review processes and internal channels for raising concerns.

OpenAI's new monitorability metrics

And on the research side of safety, OpenAI's alignment team published two new ways to measure how well we can monitor AI systems. One compares models only on the exact same examples, removing a hidden bias. The other estimates how risky a deployed system really is once monitors start blocking outputs and the model retries. The message: a single score can mislead, so report several together.

Pichai on frontier AI and compute

Staying with Google, in a story we've been following, Sundar Pichai doubled down on frontier ambitions in a new interview. He says the Gemini 4 Argon model reflects that strategy, and that Google will give up some short-term cloud revenue to feed frontier research. He described AI as increasingly helping build itself, but said Google is far from a fully automated loop and needs stronger safety monitoring. He also stressed world models for robotics, and admitted compute shortages are real.

Google and Meta launch TorchTPU

Related on the hardware front, Google and Meta teamed up on TorchTPU, an open-source project that lets PyTorch run natively on Google's TPUs with minimal code changes. Earlier integrations were clunky, so this could make TPUs a genuine alternative for the huge PyTorch community. It's notable to see two rivals cooperating here.

AI agents map brain CRISPR data

Now to science. OpenAI, working with PerturbAI and Scripps Research, says thousands of AI agents analyzed a whole-brain CRISPR dataset covering millions of mouse brain cells in just 32 minutes. Each agent took a slice of the problem, and together they built an atlas of genes, cell effects and possible mechanisms that matched expert judgment better than a single agent. Leads include a possible link between cholesterol regulation and neurodegeneration. In a related move, OpenAI acquired Orchestra Bio, a platform for coordinating drug development, to pair with its Rosalind life sciences tools. The goal is agents that help push discoveries through the slow, fragmented path to actual medicines.

OpenAI acquires Orchestra Bio

Consumer AI now. Meta's companion app Muse hit 10 million U.S. downloads in 40 days, putting it in rare company alongside mostly blockbuster games and Disney Plus. Meta's reach across Facebook and Instagram, plus heavy ad spending, clearly helped, and it reportedly retains users better than ChatGPT, Claude and Grok. Meanwhile, ChatGPT added Locked Chats, letting you protect individual conversations with a PIN or Face ID. Handy on shared devices, but OpenAI is clear this isn't end-to-end encryption and doesn't change data retention.

Meta's Muse app growth surge

For the infrastructure crowd, in an update to an ongoing story, DeepSeek open-sourced DeepInfer, the serving system behind DeepSeek V4.1 Flash. Its standout idea is keeping cached work for more than 72 hours, so repeated prompts don't get recomputed, cutting cost and latency. Separately, the open-source llm-d project hit version 1.0, aiming to make running large models on Kubernetes production-ready across many kinds of chips.

ChatGPT adds Locked Chats

Agents are getting sturdier plumbing too. Vercel released durable-harness, open-source code that lets coding agents like Claude Code and Codex survive crashes and dropped connections without redoing finished work. Ramp detailed Laminar, its in-house agent platform, which it says was first built in six weeks entirely by agents and now runs all its agentic products. And Databricks introduced MemCurrent, a memory system that helps agents remember useful lessons without drowning in stale history, reportedly tripling pass rates on internal tests. Databricks also published advice on reining in exploding AI coding bills, arguing smart routing to cheaper models and trimming wasted tokens can save a lot without hurting quality.

DeepSeek and llm-d serving stacks

In voice, Mistral launched Voxtral 3, combining transcription, expressive speech and an agent toolkit for multilingual voice assistants. Its pitch to enterprises is control, with an EU-hosted service or self-deployment. And on small models, XHToken released Spark-X2.6, a 4-billion-parameter model for on-device use, with adjustable thinking levels and an open Apache license.

Durable infrastructure for AI agents

Finally, a thoughtful essay on AI and hardware. The author argues AI may soon beat humans at pure design, but the real world is messy: measurements are imperfect, testing is slow, and manufacturing doesn't forgive. Superhuman design, in other words, isn't the same as superhuman engineering.

That's it for today's episode of The Automated Daily, AI News edition. If you run anything on PHP, now's a good time to patch. Links to all the stories we covered can be found in the episode notes. I'm TrendTeller, thanks for listening, and see you tomorrow.

More from AI News