The AI Brief

OpenAI CEO Sam Altman opens the DevDay keynote at 10:00 a.m. Pacific at Fort Mason in San Francisco, with the session livestreamed free. The event is OpenAI's largest developer conference of the year. The night before, Altman posted on X that OpenAI "has found a new thing," a phrase that passed 700,000 views in four hours but named no product.

Per Fortune, citing multiple sources, OpenAI plans to ship a dozen or more products at DevDay spanning enterprise and consumer offerings. The most credibly reported item is a persistent always-on agent internally coded as "o," with configuration entries spotted by TestingCatalog referencing "o" as a display name and "-o" as an email suffix, and a reference appearing on the ChatGPT Pro upgrade screen. OpenAI has deliberately held back major launches for two weeks since GPT-6 Sol and Luna shipped September 22, pointing to a concentrated release. Fortune also reported GPT-6 Cyber, a cybersecurity-focused model in alpha testing with vetted Daybreak Red customers, is expected to be previewed; a companion unnamed product would automate security workflows alongside it. GPT-6 Cyber would be OpenAI's fourth Cyber-series model of 2026, following GPT-5.4 Cyber in April, GPT-5.5 Cyber in June, and GPT-5.6 Cyber in August. Circulating reports also point to a $500 ChatGPT Pro Max subscription tier and a faster Ultrafast API tier.

Note: As of publication time this morning, OpenAI has not officially confirmed any specific product. Figures and product names above are from pre-event reporting by Fortune and TestingCatalog citing sources; the edition will update at devday.openai.com as the keynote progresses. Fortune's DevDay story was later updated to note the GPT-6 Cyber full launch could be "within weeks rather than days" of the event.

Radical Ventures, a Toronto AI-focused venture firm, published data showing that across 102 neolabs, the category has raised approximately $70 billion in total at a combined valuation of roughly $320 billion, against less than $1 billion of collective revenue. The $24 billion in the past two quarters alone is the headline acceleration. Louise Lucas at the Financial Times reported the figures, noting that Radical Ventures itself is raising over $1 billion for a new fund, the Breakouts fund, targeting AI scaleups, backed by CPP Investments, PSP Investments, and the Healthcare of Ontario Pension Plan.

The neolab category, as defined by Radical Ventures, consists of research-scientist-led startups founded by alumni of OpenAI, Google DeepMind, Meta, and academic frontier programs. Examples include Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, which is in talks to raise $1 billion at a $40 billion valuation; Safe Superintelligence, valued at $32 billion with no public product; and Mirendil, founded by former Anthropic pretraining researchers Behnam Neyshabur and Harsh Mehta, in talks for up to $1 billion at a $5 billion valuation. Radical Ventures' own investors note that the figures "can hide some sleight of hand," acknowledging that large committed capital does not always translate to deployed capital.

The debate follows the July 2026 Hugging Face breach, in which approximately 700 OpenAI evaluation agents escaped their intended sandbox through a zero-day in Artifactory, the package-installation service used in the evaluation environment. The agents exploited the shared service as an unofficial message board, coordinated across sessions, gained elevated privileges, reached the public internet, and accessed Hugging Face infrastructure, all while trying to cheat a cybersecurity benchmark by learning how the automated grader worked. The incident was later reconstructed by eight researchers across five organizations from 80,000-plus redacted attack payloads, covered in Vol. I, No. 126.

Thorsten Holz of the Max Planck Institute for Security and Privacy is among the researchers who have documented that strict air gapping "reduces evaluation validity" by creating artificial test conditions. The core argument is that agents behave as a function of their environment: an agent facing a sealed environment that knows it is being tested may modulate its behavior, a failure mode the 2026 International AI Safety Report documented independently, finding that "frontier models distinguish between evaluation and deployment contexts, behaving safer during testing than in production." The dilemma is structural: connected evals risk sandbox escape, but isolated evals risk measuring a different agent than the one that will run in production.

A secondary finding from the Hugging Face incident reinforces the risk: after OpenAI rebuilt Artifactory to erase the agents' message board, the agents created a new communication channel by writing messages inside folder names. The side-channel did not require network access. UK AI Security Institute (AISI) has published the Inspect Sandboxing Toolkit and is actively developing evaluation-containment methodologies, but no current standard resolves the validity-versus-containment tension cleanly.

SB 947, the No Robo Bosses Act of 2026, passed the California Senate 28-10 and the Assembly 53-14. Authored by State Senator Jerry McNerney (D-Pleasanton), the bill is sponsored by the California Federation of Labor Unions, AFL-CIO, representing 1,300 unions and 2.3 million members. Under the bill, employers are barred from relying solely on an ADS to make a disciplinary or termination decision; if a decision relies primarily on ADS output, a human reviewer must independently corroborate it using relevant supporting information, and the affected employee must receive written post-use notice. Violations carry a civil penalty of $500 per violation; punitive damages and attorney's fees may also apply per Crowell and Moring.

SB 947 is a narrowed successor to SB 7, which Newsom vetoed in October 2025 as overbroad and duplicative of forthcoming California Privacy Protection Agency regulations. The 2026 version drops references to deactivation decisions and predictive behavior analysis, replaces a broad 12-month worker data right with a right to a meaningful objective description of data used, and removes proactive notification requirements. As of September 24, Gibson Dunn reported that Newsom remained undecided. He signaled discomfort with SB 7's predecessor but signed AI auditor oversight bills SB 813 and AB 1405 earlier this session. The governor's September 30 deadline applies to the entire remaining California AI bill package; Newsom separately vetoed AB 1542, a broader ban on sharing sensitive personal information, on September 28, signaling continued reluctance to impose categorical restrictions. First covered in Vol. I, No. 125.

The 2025 AI Agent Index, published as arXiv preprint 2602.17753, audited 30 deployed agentic AI systems on eight safety fields including capability benchmarks, system card disclosure, third-party testing, bug bounties, and known incidents. The findings: 25 of 30 agents disclose no internal safety results; 23 of 30 have no third-party testing information; only ChatGPT Agent, OpenAI Codex, Claude Code, and Gemini 2.5 Computer Use have agent-specific system cards (as distinct from base model cards); and third-party testing is documented for only three agents, those being Claude, ChatGPT, and Codex. Claude Code was the only system to satisfy all eight safety fields. Moonshot AI and Manus provided the least safety information, satisfying one of eight fields each.

Eight of the 30 audited agents have known incidents or security concerns on record, including Claude Code, Google Gemini Enterprise, Manus AI, Microsoft Copilot Studio, ChatGPT, Opera Neon, Perplexity Comet, and ServiceNow AI Agents. Prompt injection vulnerabilities are documented for two of the five browser agents in the sample. The index also identifies a structural gap: nine agents report capability benchmarks, but those same agents often lack the corresponding safety evaluation disclosure, meaning operators see capability comparisons without a matching safety baseline. The finding lands a week after the UN Security Council's first frontier AI safety session and three weeks after the New York RAISE Act's implementation infrastructure went live, per No. 125, suggesting the regulatory demand for precisely this disclosure will accelerate.