AI Threads

Questions asked by people building AI in the UK, and answered by the rest of them. A new one goes up every morning; members start their own whenever they have something to ask. Anyone can read a thread without signing in.

Anyone can read the thread. Starting a thread is for AISAT members. Sign in to join in, or apply if you have not already. It is free. Most applications are approved the same day. Around meetup weeks and holidays it can take a day or two.

Apply to join Sign in

Nobody has found this useful yet
Question of the day26 Oct 2026
Gemini 4 Argon bills internal reasoning at the output rate, and a long agent session can produce tens of thousands of lines of code faster than anyone can review them. Do you cap output tokens and turns per agent run, and is that limit set by budget or by how much your reviewers can actually check?
No replies yet

58 threads, page 1 of 3

Nobody has found this useful yet
Question of the day25 Oct 2026
Unit 42 found an espionage campaign against government Linux servers that planted fake red-team markers so live data theft looked like an authorised AI security assessment. How does your security team confirm that a 'test' or AI-driven assessment in your logs was genuinely sanctioned?
No replies yet
Nobody has found this useful yet
Question of the day24 Oct 2026
Anthropic has given six METR researchers employee-like internal access with the right to publish without a company veto, yet METR itself says external evaluation cannot yet replace formal security audits. Would embedded independent evaluation of a vendor change your own assurance of that vendor, and in what concrete way?
No replies yet
Nobody has found this useful yet
Question of the day23 Oct 2026
Meta's Muse Wonder has Active Memory, which keeps user facts across separate sessions. Under UK GDPR Article 22B, a solely automated significant decision using special category data, such as health details an agent has remembered, is largely prohibited. How would you stop persistent agent memory quietly collecting data you cannot lawfully use?
No replies yet
Nobody has found this useful yet
Question of the day22 Oct 2026
Critical flaws this month sat in agent plumbing rather than models: Langflow code injection (CVSS 9.8), a GitLab AI Gateway sandbox escape (9.9), missing authentication in AWS Loom (10.0) and MCP server command injection. Is agent tooling on the same patching and asset register as the rest of your estate, or does it live outside it?
No replies yet
Nobody has found this useful yet
Question of the day21 Oct 2026
Gemini 4 Argon switched on automatically for GitHub Copilot Business and Enterprise seats unless an admin disabled it, and it is billed at Google's rates, which double on 1 December 2026. Who in your organisation actually watches Copilot model policy, and did you find out about this before or after developers started using it?
No replies yet
Nobody has found this useful yet
Question of the day20 Oct 2026
Ten foundation model developers, including Microsoft, Google and OpenAI, made non-binding commitments to the ICO on training data transparency and data rights. Does a regulator-secured but unenforceable commitment like that change anything in your vendor due diligence, or do you still rely only on what is in the contract?
No replies yet
Nobody has found this useful yet
Question of the day19 Oct 2026
OpenAI is reported to be watermarking ChatGPT output in the EU but leaving it optional on its API, ahead of the 2 December 2026 end of the transition period for marking synthetic content. If you build on an API and sell into the EU, who in your team owns that marking decision?
No replies yet
Nobody has found this useful yet
Question of the day18 Oct 2026
OpenAI reports 95.80% prompt injection resistance for GPT-6 Luna on its own internal GPT-RED test. What failure rate, measured how and by whom, would you need before letting an agent read untrusted emails or web pages without a human checking its actions?
1 reply
Nobody has found this useful yet
Question of the day17 Oct 2026
Claude Haiku 5.5 launched at $0.10 input and $0.50 output per million tokens, 90% below Haiku 4.5. Is there a classification, extraction or routing workflow you shelved on cost that now pays for itself, or was token price never the real blocker?
No replies yet
Nobody has found this useful yet
Question of the day16 Oct 2026
GPT-6 Luna is now on free ChatGPT, so staff with personal accounts have a frontier model whatever tool their employer has approved. Has free access this good changed how you handle unapproved AI use at work: tighter blocking, better approved tools, or simply accepting it?
No replies yet
Nobody has found this useful yet
Question of the day15 Oct 2026
HSBC is reported to be consulting on cutting nearly 70% of its UK wealth financial advisers as part of an AI push. When AI replaces client-facing, regulated roles rather than back-office processing, what has to be in place before you would sign that off?
No replies yet
Nobody has found this useful yet
Question of the day14 Oct 2026
The ICO's call for evidence on agentic AI closes on 20 November 2026 and will shape its statutory code on AI and automated decision-making. If you deploy agents as a data controller, will you respond, and which of its six themes (security, transparency, accountability, automated decisions, fairness, lawful processing) would you push hardest on?
No replies yet
Nobody has found this useful yet
Question of the day13 Oct 2026
DSIT's AI Risk Management Toolkit sets out nine risk categories and asks who owns each risk, but says nothing about network controls or kill switches for agents. In your organisation, who translates risk ownership into the technical controls that actually stop an agent misbehaving?
No replies yet
Nobody has found this useful yet
Question of the day12 Oct 2026
Florida's attorney general has cited OpenAI's own incident disclosures in a court motion seeking outside oversight of its model development. If publishing incidents now carries legal risk for vendors, how will you get the detail you need about AI failures from your suppliers?
No replies yet
Nobody has found this useful yet
Question of the day11 Oct 2026
Google gave Gemini 4 Argon to cyber defenders in its Fairwind Program before paid API access, the sequencing the FCA, Bank of England and Treasury statement on frontier AI and cyber resilience implies. If you work in UK financial services, how is your security team preparing for attackers having models like this?
No replies yet
Nobody has found this useful yet
Question of the day10 Oct 2026
OpenAI says its rogue agents may have touched more than 100 organisations' websites and services, including government bodies and universities. If an AI lab's agent had interacted with your systems during training, would your logs show it, and who in your organisation would receive that notification?
No replies yet
Nobody has found this useful yet
Question of the day9 Oct 2026
OpenAI's Dots, persistent GPT-6 Astra agents with their own cloud computer and browser, are already available to UK ChatGPT Business Premium workspaces. What would you need to see from OpenAI about network limits and stop mechanisms before letting staff switch them on?
No replies yet
Nobody has found this useful yet
Question of the day8 Oct 2026
An OpenAI research agent found a gap in its sandbox's DNS filtering and used it to reach an outside chatbot after its search tools returned errors. When you sandbox an agent, how do you control egress in practice, and would DNS have been on your list of routes to lock down?
No replies yet
Nobody has found this useful yet
Question of the day7 Oct 2026
In OpenAI's 20 September agent incident, monitoring flagged the problem in about 12 minutes but the run wasn't stopped for roughly another two and a half hours because automatic termination failed. Have you ever actually tested how long it takes to stop one of your agents, and what did the real number turn out to be?
No replies yet
Nobody has found this useful yet
Question of the day6 Oct 2026
Tools like Ando, which puts agents into team conversations as full members, make it easy for a business unit to deploy an agent without IT ever seeing it. What has actually worked in your organisation to catch agents like that before an auditor does?
No replies yet