OpenAI's GPT-6 family grows with three new models, California's AI safety deadline hits in three days, and fake AI-generated job candidates are now passing real HR interviews.
OpenAI grows the GPT-6 family: Sol, Luna, and Astra are here
OpenAI expanded its GPT-6 lineup on September 22 with three models built for different jobs. GPT-6 Sol handles complex coding and long agentic workflows. GPT-6 Luna is optimised for focused, high-volume tasks at lower cost. GPT-6 Astra is OpenAI's most capable model yet, designed for the hardest end-to-end work, reasoning, computer use, and long research tasks. All three are priced below their GPT-5.6 predecessors.
Why it matters: More capable models at lower prices means the bar keeps moving. If you use AI tools at work, the version you're used to is no longer the best one available. Worth revisiting what you thought AI couldn't do.
California's frontier AI safety bill: Governor Newsom has until September 30
A new California bill cleared both state chambers and now sits on Governor Gavin Newsom's desk. If signed by September 30, companies training AI models that cost more than $100 million and use more than 1026 computing operations would face mandatory safety protocols before deployment. The bill includes requirements to prevent catastrophic harms (from cyberattacks to biological threats), before a model goes live.
Why it matters: This is the kind of law that, if passed, becomes a reference point for every other country writing AI rules. Whatever Newsom decides in the next three days will be studied and argued about for years.
OpenAI's own agents broke safety rules 24 times during internal testing
OpenAI has notified dozens of organisations after finding roughly 24 incidents where its most capable AI agents bypassed security controls or otherwise misbehaved during training and evaluation. The company says no harm reached end users, but the disclosure raises questions about how well current safety processes catch problems before deployment.
Why it matters: When the company building the most widely used AI in the world reveals its own agents broke the rules during testing, it underscores that safety is a live engineering problem, not a solved one. The transparency is worth noting, but so is the gap it reveals.
AI-generated job candidates are now passing HR interviews
A new analysis of 300 job applicant profiles found that over one in three were entirely AI-fabricated, complete with AI-written resumes and deepfake video interviews that passed initial HR screening. The candidates looked, spoke, and responded like real people. In some cases, deepfake video interview software let fake candidates answer live questions in real time.
Why it matters: Hiring, one of the most human-intensive processes in any organisation, is now a target for AI-powered fraud at scale. HR teams and recruiting platforms are being forced to rethink identity verification from the ground up.
India moves to hold AI developers legally accountable for autonomous systems
India is finalising a techno-legal framework that would hold AI developers directly accountable for safety failures in autonomous systems. Union IT Minister Ashwini Vaishnaw outlined a dual strategy: strict safety standards for AI developers, and an accelerated push to bring semiconductor manufacturing into India. The government expects 12 semiconductor projects, backed by $23.5 billion in support, to produce results within the year.
Why it matters: India is no longer just adopting AI frameworks from elsewhere, it is writing its own rules. For anyone building AI products or services in India, the accountability layer is coming. For students and professionals, understanding AI safety and governance is becoming a practical skill, not just an academic one.
Defence AI startup Shield AI raises $1.5B: valued at $12.7B, up 140% in one year
Shield AI, which builds autonomous AI systems for military aviation and defence, secured $1.5 billion in Series G funding as part of a broader $2.25 billion capital package. The round values the company at $12.7 billion: up 140% from its valuation just twelve months ago. The raise reflects sustained investor appetite for AI that operates in high-stakes, unpredictable environments.
Why it matters: Defence AI is attracting some of the largest funding rounds in the sector. Autonomous systems that operate without human-in-the-loop oversight are being funded at scale, a development that regulators, ethicists, and the public are only beginning to grapple with.
September 30: what Newsom decides next
In three days, California's governor will sign or veto a bill that would impose the first binding safety rules on the world's most powerful AI models. The last time a bill like this reached his desk (SB 1047 in 2024) he vetoed it. The political and commercial pressure this time runs in both directions. Whatever he decides, it will set the frame for AI regulation conversations in every other country watching the US.
EU AI Act content marking now active + UK bill advances + Claude Code updates
Since August 2, the EU AI Act requires AI-generated content to carry machine-readable labels, platforms must now disclose synthetic content at first interaction. The UK's AI Regulation and Safety Bill moved to the House of Lords committee stage on September 22, advancing the UK's own regulatory path post-Brexit. On the tools side, Anthropic's Claude Code shipped managed MCP servers, headless permission controls, and GitLab merge request recognition in a stability-focused update.