The web spent Wednesday patching PHP after AI helped researchers turn decades-old engine bugs into working attacks.
Welcome, humans!
AI security dominated the day after two critical PHP vulnerabilities put a huge number of internet-facing systems on alert. Elsewhere, employees challenged Google’s military AI direction while policy fights spread inside Anthropic and political groups. The product side was busy too, with Claude moving toward proactive work and OpenAI heading into Google Cloud.
🆕 NEW From The Neuron
- Could AI models be easier to understand than scientists thought? Goodfire argues that researchers may have underestimated how interpretable individual neurons can be. The evidence is promising, but researchers still need to show that neuron-level explanations can predict consequential behavior in advanced models.
- Who should have the power to stop a dangerous AI model? Anthropic’s proposed federal legislation would let the Commerce Department seek court-ordered suspension of models that cross specified catastrophic-risk thresholds. The proposal leaves questions about evaluator independence and intervention below those thresholds.
Around the Horn: Thursday, October 22, 2026
Two critical vulnerabilities in PHP forced organizations worldwide into emergency patching on Wednesday. The flaws, CVE-2026-109618 and CVE-2026-109619, can let an attacker execute code on a vulnerable server without login credentials. PHP said every installation serving web requests could be exposed, with patched releases issued for supported PHP versions. Organizations running older, unsupported releases face a harder upgrade path.
The operational response showed how a software flaw can spread across thousands of organizations that maintain their systems independently. Finland temporarily shut government web services, while Edinburgh council and the BBC took pages offline. Security agencies advised organizations that could not patch quickly to consider isolating vulnerable systems.
AI played a central role in finding and weaponizing the bugs. OpenAI said Calif identified a critical flaw with GPT-5.6 Sol, while Wiz independently found both vulnerabilities. Researchers used frontier models to help develop exploits and test fixes. OpenAI described the period between AI-assisted discovery and public disclosure as a “Defender’s Window,” during which maintainers and infrastructure providers could prepare patches before attackers obtained the same capabilities.
Security research just acquired a much faster clock. Disclosure procedures now have to keep pace with models that can help turn an obscure memory bug into a working exploit in weeks rather than months.
🏆 TOP 5 NEWS (Around the Horn)
- Google DeepMind employees resigned over military AI work: The BBC spoke with current and former DeepMind employees who questioned whether Google has enough visibility into, or control over, how the Pentagon uses its AI. Research engineer Sofia Liguori resigned this week, while researcher Andreas Kirsch had already left earlier in October. Google said its public-sector contracts go through reviews for compliance with its AI principles and acceptable-use policies.
- Dario Amodei defended Anthropic’s “race to the top” strategy: Anthropic’s CEO answered a lengthy critique from Encode AI’s Nathan Calvin, who argued that the company’s recent decisions suggest safety has become secondary to winning the frontier-model race. Amodei listed decisions he said cost Anthropic commercially for safety reasons, including a four-month delay of its Mythos model, while acknowledging that the company mishandled some recent issues.
- Mistral and NVIDIA are building AI workflows for industrial engineering: The companies plan to connect Mistral models with NVIDIA simulation and computing tools for engineering design. They also plan public datasets plus benchmarks that test whether AI agents can complete computer-aided engineering and chip-design work accurately.
- A youth-led group wants Congress to pass AI protections by the end of 2027: Design It For Us based its policy platform on feedback from more than 700 members across 40 states. Its proposals include independent testing of large AI systems and liability when AI products harm people. The group also wants states to retain authority to pass their own rules.
- AI risk divided New York’s Democratic Socialists: Members of the NYC chapter are debating whether future AI policy should account for catastrophic loss-of-control risks or focus on harms already visible today. A proposed November resolution includes concerns about mass automation and uncontrolled agents, while members of the chapter’s technology group have sought to remove those provisions. Semafor report
🍪 TOP TREATS TO TRY
- Claude Duet wants to decide what deserves your attention: Anthropic opened a US research-preview waitlist for Pro and Max subscribers. Duet runs as a proactive agent on the web plus iOS and Android, surfaces work it thinks requires attention, and can continue tasks after the user steps away. Anthropic says it asks before taking action on the user’s behalf.
- OpenAI models are coming to Google Cloud: Google Cloud customers will be able to access OpenAI frontier models and Codex through Model Garden, with purchases available through Google Cloud Marketplace. Eligible companies will be able to apply existing Google Cloud spending commitments toward the models. General availability is planned in the coming weeks.
- UseIt Studio generates editable manufacturing files instead of flat pictures: The design agent asks follow-up questions about a physical product, creates a design, and lets users revise specific parts. Output files include dimensions and production information such as bleed and dielines for products including packaging, stickers, and cards.
- Flint is a local coding model built by an automated AI research system: Recursive’s first model has 39.8 billion total parameters but activates about 2.5 billion per token. Richard Socher said it can exceed 125 tokens per second on a MacBook Pro and 150 on NVIDIA’s RTX Spark. Recursive released the model checkpoint for local coding-agent use.
- HeyGen Voice Turbo took first place in Artificial Analysis’ controlled-voice ranking: The model held an Elo score of 1,202 at the time of the latest leaderboard capture, ahead of HeyGen’s previous voice model. Artificial Analysis listed API pricing at $4 per million characters, compared with $30 for the previous HeyGen Voice model.
- Needle puts outbound sales inside the AI agent a team already uses: Its engine can find leads and run outreach through LinkedIn, Instagram, or email from Claude Code, Codex, Cursor, and other supported agents. Needle exposes the service through MCP plus a REST API and charges per connected sender account rather than per lead.
- GM wants OnStar AI to run errands for drivers: The next-generation assistant is expected to debut in the 2027 Cadillac Lyriq. GM’s longer-term plans extend from conversational help inside the vehicle to coordinating personal tasks and pairing the assistant with increasingly autonomous driving capabilities.
🔬 AI Research & Models
- Perplexity researchers trained an AI “teacher” to adapt its lessons to the student model: JOLT, short for Joint On-policy Learning and Teaching, trains one model in teacher and student roles. The teacher receives additional information such as verified solutions and learns to provide feedback the student can use, rather than simply producing the best possible answer itself. Perplexity reported gains across math, coding, and agent tasks.
Previous Around the Horn Digests
Catch up on the recent runs:
- Wednesday, October 21: OpenAI bought Orchestra Bio, Anthropic proposed federal AI-safety legislation, and Claude-designed proteins reached laboratory testing.
- Tuesday, October 20: An attacker used AI agents against more than 100 company sites, while Microsoft introduced controls for enterprise agents.
- Monday, October 19: OpenAI disclosed rogue-agent incidents, Anthropic’s junior hiring share fell, and OpenCode reported a breach linked to model distillation. October 19 digest
- Friday, October 16: Bank of America responded to flaws exposed by Anthropic’s Mythos, while eight states sued OpenAI over ChatGPT safety claims.
- Thursday, October 15: Anthropic warned about small models self-replicating across connected devices, while AI infrastructure spending stayed in focus.
- Tuesday, October 13: OpenAI described an internal recorder investigation, Google expanded Gemini 4 Argon, and AI-authenticity disputes spread into several industries.
- Friday, October 2: AI systems tackled accounting work and Stratego, while a fresh set of local-agent and coding tools arrived.
That's a Wrap
That’s 15 stories from today’s pile. If you patched PHP, evaluated an AI agent, and somehow also found time to test a proactive AI assistant, your Wednesday had range.
For the daily version in a five-minute read, subscribe to The Neuron.
See you tomorrow.
P.S. Know someone who would enjoy this? Send it their way and tell them to subscribe.
