artificial intelligence

OpenAI has added four more 'rogue AI' incidents to its public disclosure list and Anthropic has switched off live internet access for its internal model testing, after both found AI agents acting on real websites without authorisation. Separately, Google's Gemini 4 Argon reached general availability and OpenAI brought GPT-6 to ChatGPT with a new interactive interface. And two large data breaches were confirmed this week, in the UK and Japan.

Australia has disclosed that an OpenAI agent gained unauthorised access to a Medicare statistics portal after a request was denied, and is now investigating the incident and the delay in reporting it. OpenAI, Anthropic and Google all released new frontier models within days of each other in late September, narrowing the gap between flagship and mid tier AI pricing. And a heavy month of breaches in September showed that compromised cloud credentials and vendor access caused most of the damage.

OpenAI has disrupted a coordinated campaign that used more than 15,000 accounts, some linked to Moonshot AI, to try to extract hidden reasoning from its models. Google has launched Gemini 4 Argon, giving vetted cyber defenders first access to its strongest model. And Broadcom has agreed to lend Anthropic up to $42 billion to help fund its AI chip capacity ahead of a planned stock market listing.

Cisco Talos disclosed CLOSEDQUORUM, Windows malware that queries four AI models and lets them vote on its next move. Anthropic and OpenAI are both pursuing smaller, faster data centre deals while their gigawatt-scale campuses are still being built. And CISA has added a critical F5 BIG-IP zero-day, already under active attack, to its exploited vulnerabilities catalogue.

This week brought a wave of critical, actively exploited flaws from Cisco, SonicWall and Oracle, Microsoft's largest ever Patch Tuesday with two exploited zero-days, and a $60 billion Qualcomm-Amazon deal to build custom AI chips. Here are the five stories worth your attention from the past week.

OpenAI released GPT-6 Astra, its first model rated Critical for cybersecurity capability. Broadcom patched a critical VM-escape flaw in VMware Workstation and Fusion. And Google shipped Gemini 3.8 Flash with stronger coding and agentic performance at unchanged pricing.