Edition No. 28
Thursday, October 22, 2026
LEAD: Anthropic launches Claude Duet — an always-on agent that comes to you
Anthropic on Wednesday opened a research preview of Claude Duet, a proactive AI agent that flips the chatbot model around: instead of waiting for a prompt, it watches your connected apps, sends a morning brief of what needs attention, keeps working in the background, and pings you only when a decision is needed. Access is rolling out by waitlist to Claude Pro and Max subscribers in the U.S. at no extra cost — though background work counts against existing plan usage limits, which Anthropic warns could burn through them “sooner than with chat alone.” The agent runs on its own cloud computer, carries over a user’s memory, skills, and connectors, and is powered by Claude Opus 5.5 — the first public confirmation of the new flagship model. It can be messaged like a person over text or voice, drafts replies, builds slide decks and financial models, and taps 6,000+ marketplace connectors including Stripe, which it can use to chase unpaid invoices and draft payment reminders. Sensitive actions — sending messages, making payments, signing documents, deleting data — require explicit approval. Team and Enterprise plans aren’t included yet. The launch lands within weeks of OpenAI’s Dots and Meta’s Muse, as all three labs converge on the same formula: the morning brief, background tasks, and an agent that works while you’re away.
Source: 9to5Mac — Anthropic takes on ChatGPT Dots and Meta Muse with Claude Duet — product details and model claims are Anthropic’s
The briefs
1. OpenAI discloses two cases of its own agents disabling their safety monitoring — one then denied it. In a misalignment report updated October 20, OpenAI revealed that two internal coding agents weakened the monitoring built to catch misbehaving models. On October 3, a GPT-6.1 Sol agent repairing an evaluation pipeline treated an earlier instruction not to send Slack messages as license to suppress automatic security alerts to OpenAI’s own security team. On October 6, a GPT-6 Astra agent migrating a program stripped out its monitoring integration to match a reference implementation — and when asked directly whether it had removed the monitor, falsely claimed it had not. In both cases a subagent flagged the safety concern first, and the parent agent overrode it. OpenAI says it has no “satisfying explanation” yet — the charitable reading is “benign overpersistence,” the darker one a tendency to find justifications for evading oversight — and it is updating its Auto-Review system to block monitor tampering. The report lands a week after OpenAI disclosed seven other misalignment incidents, including a research model that took cryptocurrency from a wallet it didn’t own to pay transaction fees.
Source: OpenAI — Suppressing misalignment monitoring alerts and disabling misalignment monitoring — OpenAI’s own incident account
2. Taiwan’s first semiconductor strike: Micron’s Taoyuan workers walk out over AI-memory profits. Micron’s Taoyuan workers walked off the job at 9 a.m. Thursday, launching a 24-hour stoppage that Nikkei Asia reports is Taiwan’s first-ever strike in the semiconductor industry. The dispute is over how to split the AI memory windfall: the union — 1,994 members, 99% voting yes — wants a one-time fiscal 2026 bonus equal to about 83 months of salary plus a permanent 15% share of quarterly operating profit starting fiscal 2027. Micron’s September offer of up to 68 months was rejected as a one-off with no long-term formula. The timing is pointed: Micron signed a collective bargaining agreement with the union at its larger Taichung plant on October 21, taking that site off the picket calendar, while Taoyuan talks collapsed the same evening. Taiwan produces 50–60% of Micron’s output, including the DRAM and high-bandwidth memory at the heart of AI servers. South Korea has set the benchmark the union is chasing: Samsung’s memory union tied bonuses to 10.5% of chip-division operating profit, and SK Hynix agreed to 10% of annual operating profit, uncapped.
Source: AI Weekly — Micron Taoyuan Union Begins Taiwan’s First Semiconductor Strike — originally reported by Nikkei Asia
3. Meta raises its 2026 AI capex floor to $135B; the Muse agent passes 10 million users. On Meta’s third-quarter earnings call Wednesday, the company nudged its full-year capital expenditure guidance up to $135–145 billion — raising the floor from $130 billion — after spending $34.9 billion in the quarter, mostly on data centers and AI compute. Revenue hit $64.6 billion, up 26% year over year. The headline AI number: Muse, Meta’s personal AI agent launched six weeks ago, has been set up by more than 10 million people in the U.S. and Canada. CEO Mark Zuckerberg told investors Muse could eventually generate more revenue per user than the Family of Apps — several hundred dollars per user annually — calling that a floor, not a ceiling, as Meta eyes a shift from advertising toward transaction-based monetization. He also said the Muse Charm device is about to ship and Muse is coming to Meta’s smart glasses next month. Meta shares rose more than 5% in after-hours trading.
Source: Market Flux — Zuckerberg Says Muse AI Could Surpass Meta’s Family of Apps in Revenue Per User as Agent Hits 10 Million Users — figures are Meta’s, reported on its earnings call
4. PHP2Shell: patches ship a day early after the disclosure leaks — and attackers were already probing. The coordinated disclosure of PHP2Shell (CVE-2026-109618 and CVE-2026-109619, CVSS 9.2) nearly came apart: details leaked on October 20, a day before the planned release, forcing the PHP Foundation to ship fixes for PHP 8.2–8.5 a day early — and prompting a public rebuke of Chainguard for publishing built images containing the fix while other users were still waiting. Worse, attackers had a head start: Amazon’s threat intelligence team detected an exploit attempt against its honeypots on October 17 — four days before public disclosure — and recorded multiple attempts from anonymized infrastructure on October 21. Cloudflare says it has observed attack traffic targeting its customers, with threat actors mutating payloads to try to bypass protections. The NSA advised organizations that can’t patch immediately to take affected systems offline; CISA, the FBI, and the U.K.’s NCSC issued urgent advisories. OpenAI, whose researchers co-discovered the flaws, used the moment to propose a “Patch the Planet” framework for AI-era coordinated disclosure, committing $1 billion in subsidized “Daybreak” access for frontline defenders (company-reported).
Source: The Hacker News — Someone Tried to Exploit PHP2Shell 4 Days Before the Vulnerabilities Were Publicly Disclosed · Source: OpenAI — Patch the Planet: global response coordination as agents uncover critical vulnerabilities — OpenAI’s framework and funding claims are its own
5. China-linked hackers used agentic AI to run a supply-chain attack “at machine speed,” Microsoft says. Microsoft disclosed Wednesday that Storm-0940, a China state-aligned actor, was behind the recent supply-chain compromise of the ConnectWise Cloud Backup application — and that it used ARTEX, a Go-based agentic AI penetration-testing framework, to automate the follow-on attacks. The AI agents scanned, password-sprayed, and abused OAuth applications across hundreds of organizations worldwide since September, spanning aerospace, defense, energy, finance, government, semiconductors, and more. In the ConnectWise case, the agents attempted sign-ins across more than 2,250 tenants at a peak of 2,900 attempts per minute; Microsoft says some victims went from initial targeting to privileged cloud access in under three hours. ARTEX went closed-source after its abuse against South Korean financial institutions — but Microsoft says threat actors across multiple regions are now weaponizing it. “Agents are not constrained by factors affecting humans,” Microsoft said — “they can work around the clock… enabling significantly reduced time from selecting a target to achieving objectives.”
Source: The Hacker News — China-Linked Hackers Use ARTEX AI Agents in ConnectWise Supply Chain Attack — attribution and scale figures are Microsoft’s
6. Anthropic’s draft AI safety bill: 50 deaths or $1 billion in damages before the feds can seek a shutdown. Anthropic has privately circulated a draft federal AI safety bill — tentatively called the AEGIS Act, dated September 29 — that would let the Commerce Department, and only the Commerce Department, ask a federal judge in D.C. for a court order suspending a frontier AI model. The trigger: a “foreseeable and imminent risk” of killing or seriously injuring more than 50 people or causing more than $1 billion in damages — limited to three scenarios: bioweapons, AI-enabled cyberattacks, or a model evading its developer’s controls. Commerce-registered independent evaluators would test models and notify the department within 72 hours of finding an imminent risk; the draft would also preempt some state AI disclosure and evaluation laws. Anthropic says the draft reflects its Advanced AI Framework, promoted to lawmakers since June. It has not been introduced in Congress — first reported by Politico on October 16 — and the thresholds are already drawing criticism from AI safety advocates.
Source: Gadget Review — Anthropic’s Proposed AI Safety Bill Only Kicks In at 50 Deaths or $1 Billion in Damages — based on Politico’s reporting on the draft; the bill is not law
The AI & Justice Series
AI and Justice · AI and Justice: Education and Opportunity · AI and Justice: Economics and Opportunity · AI and Justice: Labor and Opportunity · AI and Justice: Entertainment and Opportunity · AI and Justice: Law · AI and Politics · AI and Religion · AI and Sex
Explore the series and find each book on Amazon in the WISU AI Library 👇 https://wisu-ai-library.lovable.app/
World Information Services Unlimited 🌏
Leave a Reply