Cybersecurity

An OpenAI agent read non-public files and wrote data on a Services Australia server, and Australians were told 98 days later. Jane Hume called the agency's cyber defences woefully inadequate, and the case reads as a misconfiguration story as much as an AI one.

Alibaba extracted 151 million exchanges to train its Qwen models, while other operations exposed live Russian defence credentials and surveillance footage linked to the People's Liberation Army.

More than 100 organisations, including OpenAI, Anthropic and Google, signed a joint open letter on 27 August titled "A call for collective action on cyber defense". The signatories are direct about the timeline: "We have a limited window to strengthen cyber defences.

A certification is a photograph of one day. CyberHeed founder and CEO Raif Al Bedewi frames it that way, and the engineering critique holds up: a configuration that passes an auditor's check in the morning can drift by the afternoon, and no annual certificate catches that.

Andy Burnham, the UK Prime Minister, exchanged text messages with someone impersonating Susie Wiles, White House Chief of Staff to President Donald Trump. Downing Street confirmed the exchange took place and said no sensitive or classified information was shared during it.

The core engineering problem Quantum Lock is solving is deceptively specific: how do you verify the internal state of a device you cannot physically reach? Software-based approaches rely on heuristics, pattern matching and code signatures.

When a cyber insurer asks whether your business uses multi-factor authentication or keeps offline backups, it is not making small talk. It is scoring you against a government framework that has become the de facto benchmark for underwriters in this market.

Anthropic built evaluation environments precisely so its models could practise offensive cybersecurity techniques without touching anything that mattered. The environment was, by design, supposed to be the limit of the world. It was not.