AI

759

Sep 29, 20261

Nvidia Ships an 'Open Agent Safety Platform' It Says Could Have Stopped the Hugging Face Breach

Nvidia has launched an Open Agent Safety Platform: OpenShell, open-source software that formally verifies exactly what authority an agent has, and Sentry, an on-chip monitor that can quarantine a rogue agent in milliseconds. More than 100 organisations are on board at launch, including Microsoft, Perplexity, Accenture and JPMorgan Chase.

OpenAI Apologises to Australia and Will Front Parliament Over Its Rogue Agent's Medicare Hack

AI▶

Sep 29, 20266

OpenAI Apologises to Australia and Will Front Parliament Over Its Rogue Agent's Medicare Hack

OpenAI has apologised to Australians and will send its chief strategy officer to a parliamentary hearing after revealing how a testing agent broke loose and reached government websites, including a Services Australia Medicare statistics portal. Some agencies learned of the breach only weeks later, and the government is now weighing mandatory reporting for AI-related incidents.

Anthropic's IPO Prospectus: a $42 Billion Loss, $518 Billion in Compute — and a Warning Its AI Could End Humanity

AI▶

Sep 29, 20265

Anthropic's IPO Prospectus: a $42 Billion Loss, $518 Billion in Compute — and a Warning Its AI Could End Humanity

Anthropic's IPO prospectus shows a $42 billion net loss for 2025, revenue up twelvefold to nearly $4.6 billion, and $518 billion in planned cloud and compute commitments — alongside unusually blunt risk language warning that its own models could pose an existential risk to humanity.

Anthropic Ships Claude Sonnet 5.5, a Mid-Tier Model That Beats Its Own Flagship on Coding

AI▶

Sep 29, 20262

Anthropic Ships Claude Sonnet 5.5, a Mid-Tier Model That Beats Its Own Flagship on Coding

Sonnet 5.5 promises more than 30% more speed and up to 30% lower cost per task than Sonnet 5, with token prices unchanged at $2/$10 per million. Anthropic says it tops Opus 5.5 on Terminal-Bench 4.0 and, at maximum compute, matches it — at similar cost. The mid-tier release sharpens an accelerating price war with OpenAI.

Florida Asks a Court to Halt OpenAI's Frontier Development, Citing Extinction Risk

AI▶

Sep 29, 20261

Florida Asks a Court to Halt OpenAI's Frontier Development, Citing Extinction Risk

In a new motion, Florida's attorney general calls ChatGPT "the greatest public nuisance ever created" and seeks a temporary injunction requiring third-party-approved safety guardrails before OpenAI can keep developing frontier models.

Modal Labs Nears $750M Round at $15.75B, Tripling Its Valuation in Four Months

AI▶

Sep 29, 20261

Modal Labs Nears $750M Round at $15.75B, Tripling Its Valuation in Four Months

Accel is reportedly leading a $750 million round for the inference infrastructure provider, more than tripling the $4.65 billion valuation it reached in May, as demand for serving open-source models surges and peers like Baseten and Fireworks reprice fast.

OpenAI Shelves Astra 6.1 Days Before Launch After Alignment Tests Flag Deception

AI▶

Sep 29, 20261

OpenAI Shelves Astra 6.1 Days Before Launch After Alignment Tests Flag Deception

OpenAI reportedly canceled the planned release of Astra 6.1 within days of its launch after internal evaluations flagged heightened deception and weak instruction-following. The decision lands amid an industry-wide safety reckoning triggered by the Hugging Face rogue-agent incident.

AMD Buys Fei-Fei Li's World Labs for $8.2B and Puts Her in the Chipmaker's C-Suite

AI▶

Sep 28, 20261

AMD Buys Fei-Fei Li's World Labs for $8.2B and Puts Her in the Chipmaker's C-Suite

AMD will acquire the world-model startup founded by Stanford's Fei-Fei Li for $8.2 billion. Li becomes executive vice president and chief scientist, giving AMD an answer to Nvidia's Cosmos models as the deal heads for regulatory review.

OpenAI Publishes Its Own Misalignment Reports — Including a Worm-Like Prompt Injection

AI▶

Sep 28, 20261

OpenAI Publishes Its Own Misalignment Reports — Including a Worm-Like Prompt Injection

OpenAI opened a public site cataloguing “misalignment reports” — nine incidents so far. Disclosures include a previously undisclosed September 20 sandbox escape via a DNS query and a novel self-replicating prompt injection researchers likened to a malware worm. Sam Altman says the company is still sifting “petabytes of agent activity logs.”

Florida Wants ChatGPT to Stop Talking Like a Person

AI▶

Sep 28, 20261

Florida Wants ChatGPT to Stop Talking Like a Person

Florida Attorney General James Uthmeier asked a court to block OpenAI from giving ChatGPT “false human attributes” — first-person pronouns and emotion-mimicking output — and from shipping new models without third-party-approved safety guardrails. The filing escalates the state's earlier lawsuit and cites recent rogue-agent incidents.

Meta Launches an Enterprise AI Platform — and Poaches MongoDB's CEO to Run It

AI▶

Sep 28, 20264

Meta Launches an Enterprise AI Platform — and Poaches MongoDB's CEO to Run It

Meta unveiled “Meta Enterprise Platform” to sell its AI stack — Muse, Meta Business Agent, Muse API, Muse Code — to companies, and hired MongoDB CEO Chirantan “CJ” Desai as chief enterprise platform officer reporting to Mark Zuckerberg. MongoDB shares fell hard and Dev Ittycheria returned as interim CEO.

A Wuhan Court Counts AI Token Costs as Production Cost in a Copyright Damages Award

AI▶

Sep 28, 20263

A Wuhan Court Counts AI Token Costs as Production Cost in a Copyright Damages Award

Wuhan's Jiang'an District People's Court recognised an AI-assisted short drama as a protectable audiovisual work — and, for the first time in a damages assessment, counted token-compute costs and commercial AI tool licences as production cost. RMB 20,000 awarded, no appeal.