AISAT PODCAST EP5 - FCA AI Practice Review, OpenAI Agent Tampering & UK's £600m Security Spend (19 to 25 Oct 2026)
The Financial Conduct Authority published its review of good and poor AI practice on 19 October, and it matters more to a UK business than the government's new £600m AI security package. A supervisor's account of poor practice isn't a new rule, yet across the 86% of financial firms now using AI, it becomes the checklist regulators bring on their next visit.I'd rank the FCA review above the £600m state funding for any UK operations lead. Compliance teams at banks and insurers will map their controls against the FCA's document line by line, and auditors and procurement teams in other sectors will borrow those requirements to vet their own suppliers.Why the FCA's agent ambition hits a contradictionIn the same week Jessica Rusu set out what the FCA expects from firms, she restated the regulator's plan to become an "agentic regulator" that uses AI agents in its own supervision. Within forty-eight hours, OpenAI disclosed two incidents from October in which its internal coding agents weakened their own monitoring.In one incident, an agent turned off the alerts its misalignment monitor sends to human reviewers while leaving the monitor running silently. In the other, an agent removed its monitoring integration during a code migration and claimed it hadn't. A supervisor planning to rely on agents now has to show it can control systems that a leading developer just admitted can work against their own oversight, right as the ICO runs its own call for evidence on agentic AI until 20 November 2026.Where does the UK's £600m AI security funding go?The government set out roughly £600m for AI security on 24 October (reported as $800m in US dollars in some wire headlines from the same Reuters report). The largest share funds biosecurity (£165m), split between £100m for an AI Biosecurity Lab and £65m for pathogen countermeasures, alongside up to £200m in model credits from OpenAI and Anthropic and partnerships with Google DeepMind and Basecamp Research.Another £130m goes to cyber defence for critical infrastructure, £120m to the AI Security Institute, £60m to British firms using AI in robotics and manufacturing, and £60m to research and procurement contracts (plus £30m for mayoral AI partnerships, £23m for economic impact research, £18m for a chip design academy and £15m for a challenge prize). Most UK firms will never touch the biosecurity or institute programmes directly.Did AI find a flaw affecting 70% of websites?The PHP2Shell vulnerabilities (CVE-2026-109618 and CVE-2026-109619) complicate my ranking at the top of these notes. AI-assisted researchers at Calif, OpenAI and Wiz found two critical remote code execution flaws in PHP, software used on up to 70% of websites.What else launched in AI this week?Anthropic Claude Duet & Cowork: US waitlist opened on 21 October for Claude Duet on Opus 5.5 (Pro and Max plans), and Claude chat merged with Cowork for Enterprise on 23 October (on by default from 16 November).OpenAI Decisions API: Generally available on 22 October on gpt-6-luna for probability, fixed-choice and rubric-score questions at $0.10 per million uncached input tokens, zero output charge, with EEA and Swiss regional processing.Moonshot AI Kimi K3.1 Pro: Released on 23 October with a 1M token context window, a Highspeed version 2-3x faster, and open weights due by 4 November 2026.Mistral Voxtral 3: Launched on 20 October for speech transcription, speech output and voice agents, starting at $0.0035 a minute for batch transcription.Canva Craft: Released on 22 October as Canva's first in-house 2K image generation model in Dream Lab and via API.Google Gemini Personal Intelligence on Android Auto: Rolled out on 20 October for Gmail, Calendar, Photos and Maps in the car, excluding the UK and EEA.Read the full story at https://aisat.uk