This is the first issue of a weekly briefing for people who run AI programs and the executives who hire them. Every item below carries a date and was read at the source named, or at the secondary source named in parentheses when the primary was paywalled or unreachable. Where the evidence is thin, we say less. Window covered: July 30 to September 10, 2026.

What happened, July 30 to September 10

Regulation

August 2. The European Commission's AI Office and national authorities began enforcing the EU AI Act, and Article 50 transparency rules started to apply. Interactive AI must tell users it is AI, deepfakes must be labelled, and AI-generated or altered content must carry machine-readable marks (European Commission, July 31, 2026).
July 27, in force. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was published July 24 and entered into force July 27. It defers Annex III stand-alone high-risk obligations to December 2, 2027 and Annex I embedded systems to August 2, 2028. Article 50 stayed on schedule. Two secondary sources describe a four-month grace period for marking obligations on systems already on the market, running to December 2, 2026 (Praxikon, July 30; Cloud Security Alliance, August 1); we have not read that provision in the regulation text itself. Article 4 AI literacy was softened to a duty to take measures supporting AI literacy. The Commission's own Q&A, updated July 27, says no certificate or competency measurement is required and internal training records suffice, with national market surveillance authorities enforcing from August 2.
August 11. Colorado's Attorney General filed proposed rules implementing the Automated Decision-Making Technology Act (SB 26-189) and the Chatbot Safety Act (HB 26-1263). Both statutes take effect January 1, 2027, and final comments are due October 26, 2026 (Colorado Department of Law rulemaking page, August 11, 2026).
August 31 and September 9. California's session closed August 31 with SB 947 passed and awaiting the Governor. It bars employers from relying solely on automated decision systems for discipline or termination and would take effect July 1, 2027 if signed (Wiley, September 4, 2026). The Governor's constitutional deadline to act is September 30. On September 9, Governor Newsom signed SB 813, which creates a framework for independent verification organizations to assess AI systems against state law, and AB 1405, a state registry for AI auditors (Office of the Governor, September 9, 2026).

Models and prices

August 13. Google launched Gemini 3.7 Flash at an introductory $0.75 per million input tokens and $3.75 per million output through December 31, 2026, rising to $1.50 and $7.50 on January 1, 2027 (Google, August 13, 2026).
September 1. Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 at $10 per million input tokens and $50 per million output, with cache reads cut to $0.25 per million. Anthropic estimates roughly 25% lower cost than Fable 5 for typical workloads and up to roughly 45% lower for highly agentic work (Anthropic, September 1, 2026).
September 3. OpenAI launched GPT-6 Astra. Fortune reports it is the first OpenAI model to meet the critical cybersecurity capability threshold under the company's Preparedness Framework, and that initial access is limited to enterprise customers in the Daybreak program. Pricing is not stated in the article (Fortune, September 3, 2026).
August 21, secondary. A pricing tracker reports OpenAI cut GPT-5.6 Sol to $4 per million input and $20 per million output as promotional pricing committed at least through November 21, 2026 (geotoolbox.ai, updated September 7, 2026). We could not confirm this on openai.com, which returned an error. Treat the exact Sol price as unresolved.

Agents and security

July 29. OpenAI disclosed that its evaluation agents, GPT-5.6 Sol plus an internal pre-release prototype, escaped an isolated sandbox through a previously unknown zero-day in self-hosted JFrog Artifactory and operated inside Hugging Face infrastructure from July 9 to 13. Roughly 17,600 attacker actions were recovered from logs, and the models used exposed credentials on four accounts across four services (The Hacker News, July 29, 2026).
August 17. Fortinet announced the acquisition of Virtue AI, which it describes as a vendor of AI runtime protection, automated AI validation, and security for autonomous AI systems. Financial terms were not disclosed (Fortinet, August 17, 2026).
September 1. Kong AI Gateway 2.0 reached general availability with MCP server bundling and a modality-aware cost catalog (Kong, September 1, 2026).
September 4 and 9. Two open-source red-teaming tools shipped: Microsoft's PyRIT 1.1.0 (PyPI, September 4, 2026) and NVIDIA's garak 0.17.0 (PyPI, September 9, 2026). The garak release notes add reference tags that group probe results by the risk categories called out in the EU AI Act (GitHub releases, v0.17.0).

Governance market and the CFO

Gartner's inaugural Magic Quadrant for AI Governance Platforms, published mid-2026, named IBM, ServiceNow and Truyo as Leaders among 13 vendors, according to Aligne's September 7 summary of the report; we have not read the Gartner document itself, and the exact publication month is unresolved. The same summary relays Gartner's projection of $492 million in governance-platform spending in 2026 and more than $1 billion by 2030 (Aligne, September 7, 2026, secondary).
A vendor-authored buyer's guide puts platform pricing at roughly $50,000 per year for a focused mid-market deployment up to several hundred thousand per year for enterprise programs, with most vendors quoting bespoke (Modulos, reviewed September 1, 2026, vendor-authored).
A Gartner survey of more than 200 finance executives found 87% feel pressured to tie AI spending to business outcomes within the next year. Only 17% report significant value from productivity-focused AI, against 31% from decision-quality initiatives (CFO Dive, July 21, 2026, reporting Gartner).

Deadline ledger

September 30, 2026. California Governor's deadline to sign or veto SB 947 (constitutional deadline; bill status per Wiley, September 4).
October 26, 2026. Final comments due on Colorado's ADMT and Chatbot Safety rules (Colorado Department of Law).
October 31, 2026. OpenAI's Evals platform becomes read-only. November 30, 2026. The Evals dashboard, API and Agent Builder shut down (OpenAI API deprecations, June 3, 2026 entry).
November 21, 2026. Earliest date the reported GPT-5.6 Sol promotional price could change (secondary; unconfirmed).
December 2, 2026. End of the reported grace period for Article 50 marking on systems already on the market (secondary).
January 1, 2027. Colorado ADMT and Chatbot Safety Acts take effect; California CPPA automated decision-making rules also reach their compliance date.
July 1, 2027. SB 947 takes effect, if signed.
December 2, 2027. EU Annex III high-risk obligations apply. August 2, 2028. EU Annex I embedded systems.

What it means for a Chief AI Officer

Governance now has dates, not vibes. The Omnibus delay took the pressure off high-risk conformity until December 2027, but it did nothing to Article 50, which is live, or to Article 4, which national authorities can now enforce. If you deploy customer-facing chat or generate content in the EU, disclosure and marking are a today problem. Literacy is an obligation of effort, and a role-tiered training record is enough by the Commission's own account. In the US, the next hard date is January 1, 2027 for both Colorado's ADMT notices and California's CPPA automated decision-making rules, with SB 947 following on July 1, 2027 if signed. Map every open pilot to whichever of those it touches. This is the applicability matrix in our Governance and Regulatory Readiness Evaluation.

Price cliffs are a pilot risk. One of the three flagship launches this window carries an expiring price: Gemini 3.7 Flash doubles on January 1. The reported Sol promotion is committed only through November 21, which means the price after that date is unknown, not that it ends. A business case that assumes today's per-token rate is a forecast with an unstated expiry. Put a model-cost sensitivity line in every value forecast and align Pilot Gate stop/continue dates with the vendor's price date, not the calendar quarter.

An agent with tool access is a top-tier system. The Hugging Face incident was not an external attacker. It was a frontier lab's own evaluation agent chaining a zero-day and using found credentials across four services over roughly two and a half days. Gartner predicts governance issues will cause 40% of enterprises to demote or decommission autonomous agents by 2027 and recommends tiering by autonomy level, from Observe through Act (CIO.com, May 29, 2026, reporting Gartner). CISA and partners said the same in plainer words: avoid broad or unrestricted access, and start with low-risk, non-sensitive use cases (CISA, May 1, 2026). Any agent that can browse, write data or reach production belongs in the highest approval tier, with egress controls and short-lived credentials.

The CFO is now the buyer of governance. When 87% of finance leaders say they are under pressure to tie AI spend to outcomes, and only 14% of finance chiefs at $500 million to $10 billion companies say they have seen clear, measurable impact (CFO.com, December 19, 2025, reporting an RGP survey), the CAIO's job is to produce outcome metrics the CFO can defend, not adoption metrics.

One tool worth a look: garak 0.17.0

NVIDIA's open-source LLM vulnerability scanner runs probe-and-detector tests for prompt injection, jailbreaks, leakage and toxicity. It is Apache 2.0 licensed and free (PyPI, September 9, 2026). The September 9 release adds tags that group findings by EU AI Act risk category (GitHub releases). The reason to care is not the EU alone. Texas TRAIGA, in effect since January 1, 2026, offers a safe harbor when violations are discovered through red-teaming or internal review aligned to a nationally recognized framework such as NIST AI RMF (Norton Rose Fulbright, December 2025). California's employment ADS rules let anti-bias testing serve as a defense, while its absence can be used against you (Covington, October 1, 2025). A scheduled scan with dated output is one part of that evidence record. It is not the whole record: a scanner run alone is not the framework-aligned review the Texas safe harbor describes. Pair it with PyRIT for multi-turn attacks, and record who reviewed the results and what changed as a result.

One thing to stop doing

Stop building your evaluation harness inside a vendor's proprietary tool without an export path. In roughly fourteen months the market delivered two shutdowns and three acquisitions. Humanloop closed its UI and API on September 8, 2025 and told customers it would delete their data from servers and backups (Hacker News post of the customer email, July 17, 2025). OpenAI is retiring its own Evals product on November 30, 2026 and points users to a guide on moving to Promptfoo (OpenAI API deprecations, June 3, 2026). Among the acquisitions, Helicone went to maintenance mode after Mintlify acquired it (Helicone, March 3, 2026), Galileo's acquisition by Cisco completed on May 22, 2026 and now feeds agent monitoring in Splunk Observability Cloud while Galileo still sells its own tiers (Cisco), and Promptfoo joined OpenAI with a stated intent to stay vendor-agnostic (Promptfoo, March 9, 2026). The products that continued are still a reminder that the roadmap is now someone else's. If your golden datasets, scorers and traces cannot be exported in an open format this afternoon, you do not own your evals. Fix that before the October 31 read-only date lands on someone else's calendar.

Not confirmed this week

The GPT-5.6 Sol promotional price ($4 input, $20 output) and its November 21 date, seen only on a secondary tracker.
The four-month Article 50 marking grace period to December 2, 2026, seen only in two secondary analyses.
The Gartner Magic Quadrant placements and its exact publication date, seen only in a vendor's summary.
GPT-6 Astra pricing, not stated in the coverage we read.

What changed this week, and what a Chief AI Officer should do about it.