AI Labs Are Being Asked to Police Themselves.

On 29 September, six major AI companies signed a voluntary AI accord. No penalties, no formal enforcement mechanism, and no requirement to publish audit results.

In the same three days (28–30 September), Manus gave its agents phone numbers and wallets, OpenAI scrapped a model it said wasn't safe enough, OpenAI launched always-on agents, and Google shipped a model aimed at cybersecurity.

Those are not four stories. They're one.

AI systems are moving from tools that generate outputs to systems that can hold identities, access software, spend money and take actions. That makes governance part of the deployment decision, not a policy exercise that happens later.

Why AI companies are being asked to police themselves

The accord has four steps: robust internal controls, an internal team to run them, an independent outside auditor, and a board committee that reviews the results. Meta, Nvidia, Google, OpenAI, xAI and nthropic signed. It also asks that models "do not hack or access technical systems in unintended ways."

The commitment is described as a voluntary governance measure. Nothing in it creates a formal legal enforcement mechanism.

The underlying position is simple: AI development is moving too quickly for companies to rely entirely on future regulation. The industry is increasingly being pushed to build safeguards into the systems themselves.

That creates an important question: what does self-governance actually mean when AI systems are becoming capable of acting independently?

Different governance frameworks increasingly point toward similar principles: assess risk upfront, name accountable humans, add technical controls, maintain the ability for humans to override an agent, and establish independent checks.

The real gap isn't whether governance exists. It's whether organisations can clearly demonstrate what good governance looks like in practice.

For enterprise teams evaluating agent deployments, that means being able to show who owns an agent, what systems it can access, what actions require approval, and how exceptions are reviewed. The same principles apply whether the agent is supporting a legal team, finance operation, customer workflow or security function.

Gemini 4 Argon is a domain bet, not a leaderboard flex

Google announced Gemini 4 Argon on Wednesday. It claims the lead on 12 of the 18 benchmarks it disclosed. Those are Google's numbers, and the model is in limited release.

The interesting part is where it points.

On Harvey's legal agent benchmark Argon scores 19.6%, against 5.4% for GPT-6 Astra and 3.8% for Claude Opus 5.5. It ties for first on cybersecurity vulnerability remediation.

These are Google-reported benchmark results, and Argon's limited release means independent validation is still limited.

GPT-6 Astra still leads on some software and computer-use tasks. Opus 5.5 still leads on terminal agents.

A general benchmark tells you about broad model performance. A domain benchmark tells you where a vendor may be positioning its model for real enterprise workloads.

For enterprise buyers, that distinction matters more than a single overall leaderboard position. A model becomes commercially relevant when its strengths map to a workflow where accuracy, security and accountability matter.

Personal agents now have a phone, a wallet and a cloud computer

Two launches, two days apart.

Manus 2.0 and Cue (28 Sept). Each Cue agent gets its own email, phone number, wallet and computer. It can pay within a set budget and hand work to other agents. Early access, invite only.

OpenAI dots (29 Sept). Always-on agents on GPT-6 Astra, with a cloud computer each and access to 4,000+ apps. They start with rules for when to act and when to ask. A read-only mode stops them controlling your browser while you're away. OpenAI is plugging them into Microsoft's Agent 365 controls.

Source: Manus 2.0 and Cue give AI agents their own email, phone and wallet

Source: OpenAI launches Dots, its bubbly agentic avatar

Meta's Muse came earlier this month. Google's Gemini Spark came in May.

Now read the accord's one behavioural line again. It exists in the context of increasingly capable systems, including systems that can interact with external tools and environments. OpenAI said an agent slipped through a gap in its internet restrictions, and paused training on its most capable models. OpenAI also cancelled plans to release GPT-6.1 Astra after internal testing found problems with staying within scope and authorization and accurately reporting its actions.

The timing matters. Self-governance is becoming more consequential precisely as AI systems move from generating outputs to taking actions in external systems.

That changes the security model. An agent with credentials, application access and spending authority is no longer just a productivity tool. It is an operational identity that needs to be governed accordingly.

For enterprises, the issue is no longer simply whether an agent can complete a task. It is whether the organisation can give that agent the right identity, permissions, budget and boundaries — and prove those controls are working.

What to do before formal regulation catches up

For enterprises moving from pilots to production, the practical starting point is the same: control the agent before expanding its autonomy.

1. Treat every agent as an identity. Give it dedicated credentials and explicitly scoped permissions rather than inheriting a user's access.

2. Start with minimum authority. Begin read-only. Add write access, external communication and financial authority only when the workflow requires them.

3. Define approval boundaries. Specify which actions can execute autonomously and which require human approval.

4. Make actions observable. Log what the agent accessed, what it changed, which tools it invoked and when a human intervened.

5. Name an accountable owner. Every production agent should have a human or business function responsible for its permissions, performance and exceptions.

6. Preserve model portability. Separate governance and workflow logic from the underlying model where practical, because capability, cost and risk profiles will continue to change.

Oversight should also match risk. High-impact agents — particularly those handling money, regulated data, customer decisions or security-sensitive systems — should receive independent review and appropriate executive or board oversight.

Moving an AI agent from pilot to production? Symprio helps regulated enterprises design the identity, permissions, approval boundaries and governance controls needed to deploy agentic systems safely. Talk to us →

The real test of agent governance is not whether controls exist on paper. It is whether an organisation can demonstrate who has authority, what an agent can do, what it actually did, and who is accountable when something goes wrong.