Monday 14 September 2026

Amodei Says Slow Down. Altman Says Slow Down. Even Musk Says Slow Down. So Why Aren't They?

Dario Amodei wants the AI industry to hit the brakes — and for once, Sam Altman and Elon Musk agree with him.

Lead story

Amodei Says Slow Down. Altman Says Slow Down. Even Musk Says Slow Down. So Why Aren't They?

In a move that would have seemed unthinkable twelve months ago, three of the most powerful figures in AI — Anthropic's Dario Amodei, OpenAI's Sam Altman, and Elon Musk — all publicly agreed this weekend that the industry needs to pump the brakes. Amodei published a lengthy open letter calling on labs to "pace the frontier," arguing that safety research simply cannot keep up with the current rate of model development. Altman and Musk endorsed the call on X. Alphabet's Demis Hassabis offered what you might charitably describe as tentative support.

The trigger for Amodei's letter, at least in part, is his own company's modelling of where things are heading. Speaking separately, the Anthropic CEO warned that within six to twelve months, AI systems could be capable of coordinating swarms of agents powerful enough to "take over the entire internet." That's not a hypothetical he was dismissing — it was a warning he was using to justify the call for a slowdown.

So what does "pace the frontier" actually mean in practice? Amodei's letter is deliberately vague on specifics, which is either a strategic choice to build a broad coalition or an admission that nobody has a concrete mechanism yet. There's no proposed moratorium, no suggested benchmark for when it would be safe to speed back up. The ask seems to be: labs voluntarily agree not to deploy the next generation of models until safety evaluations have meaningfully caught up.

The political reception was predictably split. Donald Trump and House Speaker Mike Johnson pushed back hard, arguing the executives are overreacting and that any slowdown risks handing China a decisive lead in the AI race. That framing — safety vs. competitiveness — is now the central fault line in US AI policy heading into the next election cycle.

What makes this moment genuinely different from past AI safety debates is who is saying it. These aren't academics or civil society groups. They're the people building the systems, sitting on the revenue, and making the deployment decisions. When the people running the race start arguing for a speed limit, it changes the nature of the conversation — even if the actual deceleration is harder to achieve than a public letter suggests.

What to watch: Whether this coheres into a formal commitment — something like the voluntary safety pledges the White House extracted from labs in 2023, but with teeth — or whether it dissipates into positioning. The gap between Amodei's rhetoric and his company's continued model development will be scrutinised closely. Anthropic is not, as of this writing, pausing anything.

For Australia, the debate lands in an interesting spot. The federal government has taken a cautious, principles-based approach to AI governance — light-touch by design — while the EU's AI Act has set hard compliance deadlines. If the world's leading labs are now arguing that frontier AI needs formal guardrails, the case for Australia's current "wait and see" posture becomes harder to defend. The AI Safety Institute in Canberra will be watching this closely.

Also today

Passkey Phishing Is Now a Real Attack Vector — Microsoft Has the Receipts

Microsoft has detailed two live campaigns that should make security teams sit up. The first involved more than a million scam emails sent between August 3–5, with attackers impersonating CEOs using hijacked third-party email infrastructure. The second is more alarming: threat actors are using passkey-themed social engineering to trick users into handing over cloud credentials, then pivoting inside Microsoft 365 environments to exfiltrate data. Passkeys were supposed to be phishing-resistant. They still are — technically. But the attack doesn't target the passkey itself; it targets the human being prompted to approve one. The distinction matters enormously for how organisations train users and design enrolment flows.

The Hacker News ↗

Security Through Obscurity Is Dead — AI Killed It

The Register has published a sharp analysis arguing that AI-assisted reconnaissance has definitively closed the door on obscurity as a defensive strategy. The argument isn't new, but the evidence base is: automated tools can now enumerate hidden endpoints, guess internal naming conventions, and map undocumented APIs faster than any human red team. The piece is a useful read for anyone still relying on "nobody knows this endpoint exists" as part of their threat model. The implication for defenders is blunt — if your security depends on attackers not finding something, assume they already have. Document your attack surface properly, because the AI already has.

The Register ↗

Optus Outage Traced to Hardware Failure at Exchange

A hardware failure at an Optus exchange knocked out voice call services over the weekend, prompting welfare checks by police on customers who couldn't be reached. The incident is a reminder that even basic telephony remains vulnerable to single points of failure in physical infrastructure — and that outages affecting voice calls carry a different kind of risk than a slow internet connection. This is Optus's second significant network incident in the past two years. The Australian Communications and Media Authority (ACMA) has previously flagged network resilience requirements following the November 2023 outage, and it's likely this event will invite fresh regulatory scrutiny.

iTnews ↗

Perplexity Hands GPT-6 Astra the Keys to Its Production Systems

OpenAI's GPT-6 Astra is now writing internal communications, modifying software, and monitoring live production infrastructure at Perplexity — with far less human oversight than earlier models required. OpenAI is framing this as a case study in agentic AI reliability. It's also a useful data point in the ongoing debate about how quickly organisations are willing to trust AI agents with systems that actually matter. Perplexity says it "checks in much less frequently" with Astra than with its predecessors. That's either a testament to the model's capability or a preview of how accountability gets diffused when the agent is doing the work and the human is just approving it.

OpenAI Blog ↗

Sam Altman Rules Out OpenAI IPO in 2026

OpenAI CEO Sam Altman has told Fortune that listing the company this year would be "ill-advised," effectively taking a 2026 IPO off the table. Altman also addressed the Hugging Face hacking incident, discussed recursive self-improvement in AI systems, and acknowledged that building an AI beyond human control is "absolutely" possible — while insisting OpenAI would pause training before reaching that threshold. The IPO timing matters for the broader AI industry because OpenAI's public listing would set a valuation benchmark and likely trigger a wave of secondary fundraising activity. Pushing it into 2027 or beyond gives the company more time to resolve its governance structure following last year's board restructure.

The Verge ↗

Your Car Has Been Quietly Selling Your Data — and the FTC Just Noticed

The US Federal Trade Commission has issued an unprecedented five-year ban on General Motors selling customer data to consumer reporting agencies and data brokers, following years of car manufacturers quietly monetising location data, driving behaviour, and trip histories. The Verge's breakdown of how the practice evolved — and how few car buyers realised it was happening — is worth reading in full. In Australia, the Privacy Act reforms currently before parliament would tighten consent requirements for exactly this kind of secondary data use. If you drive a connected car in Australia, your insurer may already have access to data you didn't knowingly share.

The Verge ↗

AEC's Election Donation System Build Blows Past $70 Million

The Australian Electoral Commission's new election donations disclosure system has more than doubled in cost during its first nine months of construction, reaching $70 million. The system is intended to modernise how political donations are reported and made publicly searchable — a transparency priority after years of criticism about slow and opaque disclosure timelines. Cost blowouts of this scale on government IT projects are hardly unusual in Australia, but the context here is particularly pointed: a system designed to improve democratic accountability is itself becoming a case study in procurement governance failures. The AEC has not yet indicated a revised delivery timeline.

iTnews ↗

Waymo's Robotaxi Pulled Over and Called the Cops — on Its Own Passengers

A Waymo autonomous vehicle in San Francisco detected that its passengers were carrying a loaded AR-style ghost gun, pulled over, and contacted police — who arrested the two juveniles on board. Waymo confirmed the vehicle flagged a terms-of-service violation involving a firearm. The incident raises genuinely novel questions about how robotaxis mediate between passenger privacy, public safety obligations, and law enforcement. What data does the car collect? When is it obligated to share it? And who decides? Lyft is also entering the robotaxi market, per separate reporting, which means these questions will need legal answers sooner rather than later.

The Verge ↗

Automattic's Board Tried to Fire Mullenweg — and Failed

WordPress parent company Automattic has confirmed that founder Matt Mullenweg is back as chairman and CEO "with full support of the board" — which raises the obvious question of what just happened. TechCrunch reports the board attempted to oust Mullenweg before backing down. The episode follows months of turbulence at Automattic stemming from Mullenweg's very public dispute with WP Engine, which triggered a mass staff exodus and significant reputational damage to WordPress's ecosystem. Mullenweg's return with a freshly supportive board suggests the attempted ouster failed to gain enough traction — but the underlying tensions in the WordPress community haven't resolved.

TechCrunch ↗

Larry Ellison Cancels $7.5 Billion Oracle Stock Sale

Oracle founder Larry Ellison has pulled back from a previously disclosed plan to sell 50 million shares worth approximately $7.5 billion. No explanation has been given for the cancellation. Ellison is Oracle's largest individual shareholder, so a sale of this scale would have been a significant market signal — and cancelling it is arguably an equally significant one. Oracle's stock has climbed sharply over the past year on the back of AI infrastructure demand, with the company positioning its cloud division as a major beneficiary of enterprise AI workloads. Ellison walking away from a $7.5 billion exit suggests he thinks the run isn't over.

TechCrunch ↗

John Deere's Right-to-Repair Experiment: Better Than Nothing, Not Good Enough

Ars Technica put John Deere's self-repair service through its paces and came away with a nuanced verdict: the manufacturer's portal can fix some things, but farmers remain sceptical. The service requires internet connectivity, locks certain repairs behind dealer authorisation, and doesn't cover older equipment. The right-to-repair movement secured legislative wins in several US states over the past two years, and John Deere's service is partly a response to that pressure. In Australia, the ACCC has been examining right-to-repair issues across multiple industries, and agricultural equipment is specifically on its radar. The Deere model — partial access, controlled on the manufacturer's terms — is likely what Australian farmers would get under a similar arrangement.

Ars Technica ↗

Sources consulted