AI Updates

OpenAI’s public catastrophic-risk report for June 4–September 4, 2026 describes GPT-Red as restricted internal models that generate attacks—especially against prompt-injection defenses—to test and train stronger safeguards. They sit with other restricted checkpoints such as cybersecurity models and GPT-Rosalind, typically in higher-isolation L3 environments.

Google's EU Public Summary of Training Content (updated 13 October 2026) says the Gemini 4 family was trained on more than 10 trillion text tokens, over 1 billion images and over 1 million hours each of audio and video. It reveals reinforcement learning for reasoning and theorem-proving in Gemini 4 Argon, synthetic data from Gemini 3, Gemini Apps user chats, and robots.txt and Google-Extended controls, with no country breakdown.

About 2,500 Maharashtra government officials can now use Sarvam AI's Indus workspace, launched as MahaINDUS by Chief Minister Devendra Fadnavis, to draft documents, summarise files, search Government Resolutions and evaluate tenders in 22 languages. The state also signed an Enterprise Agreement with Sarvam's parent Axonwise. Officials say data stays in India and is not used to train AI models.

Headquarters Integrated Defence Staff and Sarvam AI have signed a Memorandum of Understanding to collaborate on AI applications, research, knowledge management, training and decision support across the Indian Army, Navy and Air Force. The partnership will stress secure, sovereign and scalable AI, data sovereignty and responsible AI. Sources say it runs for five years under Sarvam's defence division, Chanakya.

Anthropic said its threat intelligence team spotted the activity by watching how accounts behaved, not by reading every chat. Investigators grouped related users into internal clusters, then matched language, time zones, shared proxy servers, fake identities, virtual cards and stolen API keys. Sudden spikes in traffic — in some cases millions of exchanges a day — and the reuse of the same account pools after bans pointed to organised campaigns rather than isolated users. The company said it then

RBIH is bringing together innovators, regulators, and financial leaders to build solutions that are intelligent, inclusive, and ready for scale. For global builders, this is a rare opportunity to create technology that not only protects millions

AMLUCS 2026 continues its track record of fostering collaboration in the UK’s Cyber/AI community. Tickets are available now. For full programme details, registration and enquiries, visit https://www.amlucs.ai or contact amlucs@fnc.co.uk.

X has open-sourced the actual constants that rank its For You feed — not the architecture this time, the numbers. Read them closely and they describe a platform that stopped caring about likes a long time ago.

AISI's incident report sets out the range configuration, the agent harness and the prompts that governed Mythos 5 and GPT-5.6 Sol — and the containment that followed

Hyderabad Cyber Crime Police register Case No. 502/2026 against Instagram users and the Head of Meta in India after AI-morphed videos show the Prime Minister in a derogatory manner

Meta's invisible watermark system remains in place "even when cropped, compressed, resized, or a screenshot is taken".

Security researchers have identified a new Android malware that represents a clear step up in mobile threats. According to ESET’s H1 2026 Threat Report, the malicious app called PromptSpy can examine what appears on a user’s screen and automatically adjust its actions to fit different phones, settings, and languages.