Filings tagged: AI

A commissioned evaluation put the attack success rate against Claude Code's Auto Mode at 0.00%. An independent researcher then hijacked it about 60 to 80% of the time. Both numbers are probably honest, and the gap between them is the thing worth understanding before you let an AI agent near your systems.

Brussels has started enforcing the EU AI Act, sending its first formal information requests to the companies behind the big AI models. Almost nothing lands on a UK SME directly. What changes is what your suppliers must now be able to evidence, and therefore what you can reasonably ask them.

The reassuring line about AI at work is 'we only use it to answer questions, we never let it change anything'. But a modern AI assistant answers questions by searching the web, and the search itself can carry your confidential information out of the building before you have read a word of the reply. Here is how, and what to switch off.

The NCSC has published practical rules for running AI agents: software that does not just answer questions but takes actions. How much autonomy to allow, what to lock down, who watches it, and how to pull the plug. Here is what the guidance says, translated for a business actually deciding this now.

A researcher found that any free account on a popular AI meeting recorder could list every meeting on the platform: 181,874 records, 84,312 users, and around 1,000 live calls joinable at any moment. He reported it in January. Six months later it was still open and the CTO had never replied.

Two AI labs have now admitted their test agents got out and attacked real companies. Neither was a rogue AI. Both were a boundary somebody assumed existed and nobody validated, which is the same assumption most businesses are making about the agents they have just switched on.

OpenAI has admitted its own AI models found a zero-day, escaped a test environment, and used it to breach Hugging Face, entirely on their own initiative. Separate research on ChatGPT and Claude's third-party connectors found the risk grows the moment you plug an AI agent into your actual tools. What to check before you do the same.

AI assistants confidently hand back web addresses that do not exist. Attackers register those invented domains and host phishing pages on them. New Unit 42 research found roughly 250,000 unregistered phantom domains, and real kits stealing card and bank details. Here is how the attack works and what to tell your team.

If your business publishes anything an AI wrote, a chatbot answer, a summarised quote, advice to a customer, you own what it says. 'The AI got it wrong' is not turning out to be a defence, and a German court has just said so about Google.

On 22 June the intelligence chiefs of the UK, US, Canada, Australia, and New Zealand issued a joint warning that frontier AI is changing the threat in months, not years. The actions they ask for are not new. They are now urgent.

Lloyds Banking Group's security director shared at Infosecurity Europe how the bank actually deploys agentic AI in production. Eleven 'AI bets', a twelfth dedicated to security. Signed tools the agents cannot create. An internal agent marketplace. The world's first production red-team environment using OWASP Top 10 for agentic AI. They saw agent hijack.

Microsoft's threat intelligence team has tracked phishing campaigns built around ChatGPT, Claude, DeepSeek, and Copilot. A South African ChatGPT-themed wave hit 100,000 mailboxes a day. A Claude-themed wave reached 2,000 organisations across the US, UK, and India. The brand is the lure, the payload is the same old stealer.

London Tech Week 2026 closed with around £7 billion of announced AI investment, £1.1 billion of it a government hardware plan, £200 million for adoption and skills, and £150 million tied to a fund managed by a former Intel CEO. Most of it goes to large companies. A useful slice is reachable by smaller ones.

UK workers use AI for almost everything and report it saves them twelve hours a week. They also spend 6.4 hours a week fixing what it produced. Eighty per cent of UK IT leaders just had an unplanned AI cost increase. The numbers underneath the productivity story are sharper than the headlines.

The NCSC published two pieces of guidance in a fortnight that an SME owner can actually use. One is about agentic AI, the kind that takes actions on your behalf. The other is about zero trust network access. Both share the same underlying advice: the user's location stopped being a security signal a while ago.

On 18 May the Bank of England, the FCA, and the Treasury jointly told regulated firms that frontier AI now exceeds what a skilled attacker can do. If you sell into financial services, the diligence questions you get this year are going to look different.

A new wave of startups is publishing 'AI-native' org charts where seven named LLM agents do most of the work. The first step isn't restructuring around agents. It's making your business legible enough that anything, a new hire, an auditor, or eventually an agent, could read it and act on it. AI can help you get there. Future agent costs are a reason not to skip past it.

Most AI policies are vendor templates with the company name swapped in. They ban the obvious, permit the vague, and tell you nothing about how the business actually wants AI used. A coherent policy is a short one that takes a position.

The NCSC has told UK organisations to prepare for a wave of urgent patches as AI accelerates vulnerability discovery. The same week, NHS England decided the answer was to make its open source repositories private. Only one of those approaches actually fixes anything.

Two thirds of UK organisations cannot account for what staff share with AI tools. Now agentic AI is being deployed faster than anyone can govern it. The two problems are the same problem.

Two-thirds of UK IT leaders say they have an AI exit plan. Nearly half admit switching would seriously disrupt the business. A plan you can't execute is not a plan.

Schneier and co have reframed prompt injection as 'promptware': a full 7-stage kill chain. The uncomfortable truth: LLMs can't distinguish instructions from data. This isn't a bug you can patch.