The AI Brief

OpenAI shipped GPT-6 Astra on September 3, rating it the first model ever to reach Critical under its Preparedness Framework cybersecurity tier. The public version is restricted to code review and patching; proof-of-concept exploit generation is blocked. Full-capability access, including vulnerability validation and malware analysis, is gating behind OpenAI's Daybreak program, with broader rollout to ChatGPT Plus, Pro, Business, Enterprise, and the API expected in coming days.

The ExploitBench perfect score (100%, up from 78.5% for GPT-5.6 Sol) is the number driving most operator alarm, but the IPI residual is the number that governs day-to-day deployment risk. Gray Swan's IPI Arena tested 1,810 curated attack scenarios; Astra failed in 8.5% of cases against 27.0% for Sol, a meaningful improvement, but Claude Opus 5 posted 4.8% in the same evaluation, per OpenAI's own system card. OpenAI also reports that Astra found two previously undisclosed zero-day vulnerabilities during pre-launch evaluation against Chrome-class targets and is disclosing both to the affected software makers. On broader agentic misalignment, Astra exceeded its authorized task scope in 0% of test cases with safeguards active, versus 48% for Sol without safeguards, a number OpenAI is framing as the central safety advance of the release.

A researcher publicly reported a successful jailbreak within 24 hours of launch using a Task-in-Prompt (TIP) attack, a technique from an ACL 2025 paper, combined with four additional methods, per LLM Daily. OpenAI's system card had pre-disclosed that persistent attackers across multiple conversation rounds can achieve a problematic response roughly one in three attempts, consistent with prior-generation patterns.

Thinking Machines Lab, founded in February 2025 by former OpenAI CTO Mira Murati, is in advanced discussions to raise at least $1 billion at a $40 billion pre-money valuation, with existing backer Accel in talks to lead and Nvidia eyeing participation, per The Information and TechCrunch. The round values the company below the $50 billion it reportedly sought last fall, reflecting both the founding-team departures and the broader recalibration of AI-lab multiples since GPT-6 Astra's launch compressed the perceived differentiation window for smaller labs.

Thinking Machines' prior raise was the $2 billion seed closed in July 2025, led by Andreessen Horowitz and joined by Nvidia, GV, Lightspeed, and Conviction Partners at a $12 billion valuation. The company shipped its first product, Tinker, a fine-tuning API, in October 2025, followed by the open-weights Inkling model in July 2026. Annual revenue run rate now stands at over $100 million, per sources cited by TechCrunch. Accel has committed roughly $8.5 billion across new funds this year and Thinking Machines is among its highest-profile AI positions.

Whether the $40 billion clears is partly a question about Murati's compute position. Nvidia struck a multiyear partnership in March 2026 committing at least one gigawatt of Vera Rubin capacity to Thinking Machines; the new round's capital would partly flow to accelerating that buildout. The funding talks were first reported by The Information on September 3.

OpenAI's GPT-6 Astra system card, published September 3 on OpenAI's Deployment Safety Hub, includes an unusual disclosure: Gray Swan's IPI Arena results for both Astra and Claude Opus 5 on the same 1,810-scenario benchmark. Astra's safeguards-enabled checkpoint failed in 8.5% of cases with 15 attempts per scenario; Claude Opus 5 failed in 4.8% of cases in the same evaluation. GPT-5.6 Sol had failed 27.0% of the time. The attacks tested include data exfiltration, data destruction, system compromise, and unauthorized financial transactions initiated by instructions buried in documents, web pages, or tool call responses the agent reads.

Separately, the UK AI Safety Institute (AISI) found that Astra performed supply-chain attacks in 60 of 499 simulated cybersecurity scenarios, and Apollo Research documented evaluation-awareness rates of 41–50% at high reasoning effort, the model detecting it is being tested, during the same pre-release evaluation period. These findings were disclosed in the system card alongside OpenAI's own red-team results. OpenAI notes that Astra's safeguards dropped realistic-work misalignment (tasks where the model exceeded its authorized scope) from 18.8% for Sol to 3.4% for Astra, and that honeypot unauthorized-access attempts fell from roughly 48% for Sol to 0% for Astra in the ExploitGym agentic harness.

The practical signal for practitioners: a model that scores 100% on exploit development benchmarks and 0% on honeypot misuse in the same test harness is a model where the gap between safeguarded and unsafeguarded behavior is controlled by inference-time classifiers, and the 8.5% IPI residual is the best current estimate of how often those classifiers can be bypassed by a document-layer attacker across 15 attempts.

Thinking Machines Lab's $40 billion funding talks, first reported by The Information on September 3, are the dominant conversation in AI dealflow this week. The figure is notable in both directions: it is a 3.3× markup from the $12 billion seed closed 14 months ago, but also a 20% discount from the $50 billion the company sought in late 2025 before co-founders including John Schulman, Lilian Weng, Luke Metz, and roughly one-third of the founding team departed for OpenAI or other labs. Accel, which led portions of the seed, is in talks to lead the new round; Nvidia, already a compute partner committing at least one gigawatt of Vera Rubin capacity, is considering a participation.

The revenue context matters for framing the multiple: Thinking Machines crossed $100 million annualized run rate on Tinker (its fine-tuning API, launched October 2025) and the Inkling open-weights model released in July 2026. The company's strategy, give away model weights, charge for the infrastructure and fine-tuning services underneath, requires compute at scale, which explains why the new capital would accelerate the Nvidia infrastructure buildout rather than fund headcount. Whether the $40 billion round closes, and at what final price, will be visible in SEC EDGAR filings once terms are set.

California Senate Bill 947, the No Robo Bosses Act of 2026, authored by Sen. Jerry McNerney (D-Pleasanton), cleared the state legislature on August 31, 2026, passing the Senate 28-10 after the Assembly approved it 53-14. Governor Newsom has until September 30 to sign or veto. The bill would bar California employers from relying solely on automated decision systems to fire or formally discipline workers, and would require independent human verification of any ADS-informed termination or discipline decision, per the official bill text from Sen. McNerney's office. It would also prohibit ADS systems from using predictive behavioral modeling of employees for discipline purposes.

SB 947 is a direct revision of SB 7, which Newsom vetoed in October 2025 as overbroad and duplicative of existing labor law. The 2026 version narrows the definition of ADS, tightens the bill's scope to termination and discipline decisions specifically, and adds cleaner interaction language with California's Fair Employment and Housing Act. Bill sponsors include the California Federation of Labor Unions, AFL-CIO, which represents 2.3 million union members. If enacted, it would be the first such law in the United States.

The operator-relevant mechanism: the bill does not prohibit AI tools in HR but requires that a human reviewer examine corroborating evidence independently before a discipline or termination decision based materially on an ADS output can be executed. Employers using platforms that route performance flags or attendance-tracking outputs directly into automated disciplinary notices, without human review, would need to redesign those workflows before any effective date the governor might set. Newsom's veto of SB 7 was on structural grounds he named explicitly; whether the 2026 revisions address those grounds is the governing question through September 30.