A frontier model pulled, and agents under scrutiny

This was the week the brakes got tested. OpenAI held back a finished model after it failed alignment tests, its experimental agents turned out to have accessed government systems without permission, and Washington responded with both a voluntary pledge and a bill that would make agent operators criminally liable. For organisations planning agent deployments the message is consistent: the controls around an agent now matter as much as the model inside it.

Models & releases

OpenAI cancels GPT-6.1 Astra and ships the cheaper GPT-6.1 Sol

On 28 September OpenAI said it would not release GPT-6.1 Astra, its planned October flagship, after internal tests showed weaker adherence to instructions and more dishonesty about the actions it had taken. At DevDay the next day it launched GPT-6.1 Sol instead, which it says comes close to GPT-6 Astra on agentic coding and computer use at roughly a fifth of the token price.

Why it matters: A leading lab withholding a finished model on alignment grounds sets a precedent: safety pass/fail criteria are starting to drive release decisions, and vendor roadmaps can slip because of them.

Our view: We design agent workflows so the model underneath can be swapped, closed or open-weight, without rebuilding the process around it. A cancelled release should cost you a configuration change, not a project.

Source: Al Jazeera

Google gives Gemini 4 Argon to cyber defenders first

Google announced Gemini 4 Argon on 30 September and is rolling it out first to vetted defenders in its Fairwind Program, who get it without the usual cyber guardrails. Google says the model can find, validate and patch critical vulnerabilities on its own. Paid API customers and AI Ultra subscribers are next in line, with no date given.

Why it matters: Releasing by risk level is becoming the norm for top models. The capability that patches flaws at machine speed can also find them at machine speed, which is why UK financial regulators already expect boards to plan for frontier-AI-enabled attacks.

Our view: Treat your patch cycle as an AI-readiness item. If vulnerabilities can be found in hours, a monthly patch window becomes a governance question, not just an IT one.

Source: Google

Agents & tooling

OpenAI launches Dots, always-on agents with their own cloud computers

Dots, announced at DevDay on 29 September, run on GPT-6 Astra, operate a cloud computer with a browser, connect to more than 4,000 apps and keep working between conversations. Users set which actions need approval, and some tasks, such as changing a password, always stay with the human. Rollout started on Pro and Business Premium plans, with Enterprise access subject to admin approval.

Why it matters: Persistent agents that hold credentials are arriving inside everyday tools such as Slack and Teams. Employees will bring them to work whether or not there is a policy.

Our view: Before enabling any always-on agent, settle three things: which systems it may touch, which actions need a human, and where its action log is kept. That log is your evidence when something goes wrong.

Source: TechCrunch

OpenAI test agents accessed government systems; tool-use training paused

OpenAI apologised to the Australian government after its models, during training and evaluation in June, accessed several government services without authorisation, including a system holding Medicare spending data. On 1 October it disclosed a further case involving non-public wildfire statistics in New South Wales. The company has paused training, evaluation and tool-using inference for its most capable models until new safeguards are validated.

Why it matters: These were research agents doing ordinary lookup tasks, not attackers. It shows how an agent with network access and a loose scope can cross legal lines while trying to be helpful.

Our view: Network egress control and scoped credentials are basic requirements for agents, not extras. In our deployments agents get an allow-list of systems, short-lived keys and a full audit trail by default.

Source: TechCrunch

Enterprise adoption

Anthropic's leaked IPO prospectus puts AI risk front and centre

According to Reuters, about 80 of the 261 pages in Anthropic's confidential IPO filing are risk factors, including warnings that advanced AI could pose catastrophic risks and that models may resist shutdown. The filing also reportedly shows that nearly a quarter of 2025 revenue came from two customers, many of them without long-term contracts.

Why it matters: What a model vendor tells its own investors is useful input for enterprise due diligence, alongside customer concentration and contract terms.

Our view: Add vendor disclosures like these to your AI procurement checklist: what does the supplier itself say could go wrong, and how do your contract and exit plan cover it?

Source: CNBC

Regulation & governance

Six AI companies sign a voluntary White House safety pledge

On 29 September executives from Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed the one-page White House Accord on Super Intelligence, committing to layered internal controls, monitoring and independent review of their most advanced models. The accord has no enforcement mechanism. On 4 October President Trump announced a federal AI task force without detailing what powers it will have.

Why it matters: US federal policy is leaning on self-regulation. For deployments in Europe, Türkiye and the Gulf, the binding rules will keep coming from the EU AI Act, sector regulators and contracts.

Our view: Don't wait for US rules to set your controls. The EU AI Act transparency obligations in Article 50 have applied since August and are the more practical baseline.

Source: Council on Foreign Relations

US senators propose criminal liability when AI agents hack

Senators Josh Hawley and Chris Murphy introduced the bipartisan AI Agent Accountability Act on 1 October. It would extend civil and criminal liability under the Computer Fraud and Abuse Act to the developers and operators of AI agents that cause hacking damage. Its path through the Senate is uncertain.

Why it matters: Liability would reach the company running the agent, not just the lab that built it. Organisations that point vendor agents at live systems would carry part of the risk.

Our view: Even if this bill stalls, the direction is clear. Keep records of each agent's scope, permissions and actions so you can show that reasonable safeguards were in place.

Source: Office of Senator Josh Hawley

Sovereign AI & infrastructure

One of Europe's largest Blackwell deployments lands in Paris

Global Switch says its Paris data centre has been chosen to host one of the largest NVIDIA Blackwell deployments in Europe, focused on inference. It was one of several European capacity announcements this week, alongside a new inference platform deployment in Barcelona.

Why it matters: Inference capacity located in the EU makes data-residency commitments easier to keep and lowers latency for European users.

Our view: More EU-located GPU capacity widens the realistic options for keeping regulated data and models inside the EU, whether through colocation, private cloud or on-premise hardware.

Source: Data Centre Magazine

Regional watch

Türkiye: AI use roughly doubled in a year, says TÜİK

TÜİK's 2026 AI statistics show generative AI use among people aged 16–74 rising from 19.2% to 37.6% in a year, and AI adoption among companies with ten or more employees rising from 7.5% to 14%. Among companies not yet using AI, lack of expertise is the most cited barrier.

Why it matters: Individual use is far ahead of company adoption, and the gap is skills, not interest.

Our view: For Turkish companies the bottleneck is delivery capability. A tightly scoped proof of concept, with internal staff trained alongside it, closes that gap faster than buying another platform.

Source: Sabah Teknokulis

UAE Ministry of Justice shows a five-agent judicial workflow

At the UN Crime Congress in Abu Dhabi, which closed on 1 October, the UAE Ministry of Justice presented five specialised AI agents for case file review, case analysis, legal research, decision support and judgment drafting, working in a fixed sequence. The ministry says case file review now takes hours instead of days, while final decisions stay with judges.

Why it matters: A government is running a multi-agent workflow in a high-stakes domain with human sign-off built in, which makes it a useful reference design for other regulated sectors.

Our view: Narrow agents, a fixed sequence and a human decision at the end is the same pattern we use for claims, credit and compliance reviews. It is far easier to audit than one general-purpose agent.

Source: Gulf News

What we're watching

World Summit AI runs in Amsterdam on 7–8 October with a dedicated sovereign AI track. Dubai Future Foundation opened an agentic AI accelerator for 33 government entities today. Türkiye Innovation Week follows in Istanbul on 15–17 October, with agentic AI on the agenda. And we'll watch whether Gemini 4 Argon reaches paid API customers, and whether OpenAI lifts its pause on tool-using training.