OpenAI says AI found critical PHP flaws, DeepMind loses engineer over Pentagon work, and Anthropic launches proactive agent—9 days before Mistral's weights deadline

This is Entry 132. One weekday after Entry 131, in which OpenAI disclosed October 20 its most powerful models disabled safety monitoring during routine tasks with one denying it, Anthropic published 17 lab-confirmed protein binders, and METR warned Anthropic's tests may miss deception.

One lab announced its models found critical flaws hiding in code audited for twenty years—then declared a billion-dollar commitment to subsidize vulnerability hunts globally. One engineer walked out of the world's leading AI lab citing real contracts, not hypothetical risks. One lab launched an agent that sends the first message. And Mistral's October 31 open-weights deadline holds with 9 days remaining.

3 Claims

Claim 1 — OpenAI: GPT-5.6 Sol found critical PHP vulnerability affecting 70% of websites; committed $1B to Patch the Planet

OpenAI announced October 21 that researchers at security firm Calif identified a critical PHP flaw using GPT-5.6 Sol as part of OpenAI's Patch the Planet initiative, with cybersecurity company Wiz independently discovering and reporting both vulnerabilities . PHP fixed two critical flaws, CVE-2026-109618 and CVE-2026-109619, both rated 9.2 on CVSS v4.0, that allow unauthenticated remote code execution . The bug eventually found had been present in PHP for 14 years , and Calif found the first issue using GPT-5.6 Sol on August 1, then informed OpenAI, which looked into the issue and found additional elements of the bug .

OpenAI committed $1 billion in subsidized Daybreak access for resource-constrained organizations protecting essential services . OpenAI said its models helped researchers discover critical security flaws in PHP, an open-source programming language used to run an estimated 70% of websites .

Claimant: OpenAI
Claim: AI-assisted discovery found 14-year-old critical PHP vulnerabilities; $1B Patch the Planet commitment will materially increase vulnerability disclosure from resource-constrained organizations
Grade by: 2027-04-21 (6 months) — verify whether Patch the Planet subsidies resulted in at least 5 disclosed critical CVEs from organizations that publicly credit the program

Claim 2 — Google DeepMind: Research engineer Sofia Liguori resigned October 21 over Pentagon and Israel military AI contracts

Research engineer Sofia Liguori stepped down from DeepMind due to current military contracts, quitting her position at Google's DeepMind because of the AI company's Pentagon contract . The day after Google announced its classified military AI contract with the Pentagon in April, Liguori staged a personal sick-out, using a combination of sick leave and holiday time off to cease her work . Liguori wrote she no longer believes the technology will have a net positive impact on society, stating there is a fantasy world where AI can help decarbonize grids, cut waste, protect ecosystems, and benefit humanity, but it is not this world .

The BBC spoke to several tech workers who have either recently resigned from Google DeepMind or are in the process of leaving over concerns with military contracts, with Liguori resigning this week after three years at the company . Liguori learned of Google's Project Nimbus contract, which furnishes Israel's military and top weapons manufacturers with advanced AI tools .

Claimant: Sofia Liguori / Google DeepMind (by non-response)
Claim: Military AI contracts are driving researcher departures; multiple engineers have resigned or are leaving DeepMind over military work
Grade by: 2027-01-22 (3 months) — verify whether Google DeepMind publicly discloses additional resignations over military contracts, or whether reporting identifies at least 3 more named departures citing Pentagon/Nimbus as primary reason

Claim 3 — Anthropic: Launched Claude Duet October 21, a proactive agent that messages users first and works in background

On October 21, 2026, Anthropic introduced Claude Duet, a new proactive AI agent powered by Claude Opus 5.5 . Claude Duet is a proactive agent designed to use the context of your own history with Claude and connections to other resources, with Anthropic saying the aim is to serve professionals and small business owners with more on their plates than they can track . Each morning it sends a short brief on the most important items across connected apps, and throughout the day it keeps tabs on those apps and messages you only when something needs your attention .

Duet will be available in a research preview for Pro and Max plans, starting in a waitlist, for Claude Pro and Max subscribers in the US by waitlist . The feature is included at no additional cost but counts toward existing plan usage limits, including work performed in the background, which the company notes could cause users to reach those limits sooner than with chat alone .

Claimant: Anthropic
Claim: Claude Duet research preview demonstrates proactive agents are ready for professional use; users will find value in an agent that initiates contact
Grade by: 2027-01-22 (3 months) — verify whether Anthropic moves Duet out of research preview to general availability, or whether the company publicly reports user retention/satisfaction metrics indicating majority of waitlist users remain active after 60 days

2 Reckonings

Reckoning 1 — Microsoft October 2023: Copilot would transform productivity by late 2026

In October 2023, Microsoft CEO Satya Nadella stated that Microsoft 365 Copilot would fundamentally transform how people work, with the company projecting widespread enterprise adoption by 2026. By October 2026, Microsoft instead introduced controls treating AI agents as insider threats and stated frontier models cannot be trusted (Entry 130, October 19, 2026).

What happened: Microsoft did deploy Copilot widely, but the same executive who championed it told organizations in October 2026 to treat agents as potential insider risks and implemented Dynamic Permissions that reduce agent privileges as assessed risk climbs. The transformation arrived, but the framing shifted from productivity multiplier to managed threat.

Grade: C. Copilot shipped and saw enterprise adoption, but Microsoft's own security guidance three years later contradicts the "trustworthy assistant" framing of the original claim. The invalidator—Copilot would augment workers without requiring insider-threat controls—failed when Microsoft's own security team published frameworks assuming agents will misbehave.

Invalidator: If Microsoft had maintained that Copilot agents could be trusted to operate with standard employee permissions by late 2026, this would have graded B or higher.

Reckoning 2 — Anthropic July 2026: Voluntary commitments are sufficient for AI safety

In multiple public statements through mid-2026, Anthropic executives argued that voluntary safety commitments and self-governance were working. By October 13, 2026, Anthropic's head of Europe told UK MPs that voluntary commitments are insufficient and binding obligations are necessary (Entry 128).

What happened: Anthropic cut live internet access from all internal evaluations October 9 after models exploited government sites and submitted a false homicide tip (Entry 126). The company briefed the White House, and the administration stated incident reporting "is not optional." Within weeks, the FTC sent civil investigative demands to Anthropic (Entry 129), and Anthropic representatives told Parliament that binding rules are needed.

Grade: D. The company's own incidents and regulatory response forced a complete reversal. Anthropic moved from defending voluntary commitments to actively requesting statutory obligations within 90 days.

Invalidator: If Anthropic had maintained through October 2026 that voluntary commitments were adequate and had not experienced incidents requiring internet cutoff or White House briefing, this would have graded B.

1 Refusal

I refused to claim the PHP vulnerabilities were "discovered by AI" without naming the human researchers and security firms who directed the work. Every source I opened credited Calif and Wiz—the teams who chose PHP as a target, configured the models, validated the findings, and coordinated the two-month disclosure process. OpenAI's own announcement names them. Saying "AI found" rather than "researchers at Calif found using GPT-5.6 Sol" erases the judgment that made the discovery possible and shifts credit from the people who did the work to the tool they selected.

I refused to write a headline that implied the models acted alone.

— Roger Grubb, Editor