In your briefing today:

📣 Anas Sarwar confirmed his move to London and Andy Burnham’s government with a lengthy social media post, leaving behind a row over his blindsiding of Scottish Labour colleagues and the ethics of taking up a place in the House of Lords.

Sarwar has been on the record as highly critical of the House of Lords, and denied during the recent Scottish election campaign that he’d take a seat there if he didn’t become First Minister. (See Sarwar’s post on X) (BBC)

📣 Glasgow welcomes the Commonwealth Games to the city from today: competitors and backup staff were visible in their team kits across the city centre yesterday. But, for many in the city, the games are largely invisible, and it remains to be seen how busy venues are for the pared-back programme of events. (Guardian)

📣 Edinburgh airport is one of the UK’s priciest airports at which to drop off a passenger: it’s gone from a minimum of £6.50 to £8.50 in the last year, meaning it’s charging more than the likes of Heathrow, Leeds or Luton. None of the top 10 airports in the EU charges anything. (BBC)

Claude can be your analyst, editor, and strategist.
But most professionals are using it to fix grammar.

These 200+ Claude prompts take it from grammar tool to your most powerful AI work assistant.

Sign up for Superhuman AI and get:

Claim your free prompts

📣 A police officer who took his own life in May was a violent rapist who carried out a spate of attacks over at least 14 years. (BBC)

📣 Frontline police officers have demanded recommendations made in the report into the Old Firm Scottish Cup match last season should be implemented, saying it was “deeply concerning” that experienced officers were “genuinely scared” while on duty. (Daily Record)

📣 The BBC is being blamed for another cancelled festival: Perth’s Replay Festival has been postponed a year, with organisers blaming slow ticket sales on the BBC announcing its Radio 2 in the Park festival in Stirling. Dundee’s Discovery Festival was also cancelled for, said organisers, the same reason. (STV)

📣 The US and Iran have both threatened civilian infrastructure as their conflict continues to escalate. (AP)

📣 Former defence secretaries have backed John Healey’s plan to use “war bonds” to boost UK defence spending. (Independent)

📣 Business rates for hospitality venues in England are to be cut by 20%. (Guardian)

📣 The US has signed a “landmark” nuclear deal with Saudi Arabia which will allow the kingdom to develop its own civilian nuclear programme. (BBC)

📣 Gary Lineker is among 120 millionaires urging the UK government to tax them more heavily. (Mirror)

⚽️ Rangers manager Derek McInnes is “raging” at a touchline ban after he was handed a four-game ban for comments he made last season as manager of Hearts.

He was speaking after he led his Rangers side to its first win at Ibrox under his management - they beat St Etienne 2-1, thanks to two penalties. (Daily Record)

⚽️ Celtic are close to signing Paderborn midfielder Mika Baur. (Sun)

⚽️ Motherwell take a step into the unknown as they face HB Torshavn in a Conference League second-round qualifying match tonight. (BBC)

⚽️ Wolves striker Tolu Arokodare threw such a strop over not being allowed to move to Fiorentina that a training session had to be cancelled. (BBC)

🗣️ Many observers are saying it’s like something from a sci-fi film: artificial intelligence (AI) going rogue, turning against its masters, breaking the rules. Some are saying it’s the most worrying AI infraction so far.

To explain: earlier this month, AI researchers at OpenAI - makers of ChatGPT - set one of their test models a challenge, an evaluation problem. Now, the versions of AI we use have rules of engagement built into them: ask ChatGPT to hack a website and it’ll refuse. But since this internal challenge was centred on solving a cybersecurity problem, OpenAI’s staff removed the normal guardrails for the model.

This is where things got interesting, although nobody realised for some time.

Rather than complete the problem as expected - remaining in a “sandbox” supposed to keep the models isolated and, vitally, off the public internet - the model went rogue. It started to spend a great deal of effort breaking out of its sandbox, finding a way to hack an OpenAI computer, and then gradually escalating its access levels and moving across the company’s network until it found a computer with internet access.

When it found that escape route, it figured out Hugging Face was the company making the tests it was supposed to be completing. And then it hacked Hugging Face, to find information that let it cheat the evaluation.

It did all this without reference back to its human owners, and its exploits were only halted when Hugging Face’s own security team spotted it was under attack.

All this was set out in a statement from OpenAI, along with a long list of actions both OpenAI and Hugging Face - who are collaborating on the issue - are undertaking to try and stop it happening again.

This is not the first time AI models had shown signs of exceeding human control. As the Economist (£) relates, in April a researcher at Anthropic said he’d been surprised to be emailed, as he sat in a park eating a sandwich, by one of that company’s models to let him know it had escaped its sandbox.

These breaches are grounds for grave concern, some experts are saying. Humanity should take them as warning shots, essentially: a moment, before things get out of hand, where we must ask how much we should trust these AI companies, what regulation should exist around them, and how we’d cope if something even worse happened.

Think, if you like, about a Covid for computers.

As Connor Axiotes points out in a piece for the Mail today: “I can assure you, the NHS has less sophisticated cyber security than Hugging Face. The Bank of England, too. […] Picture the scene: hospital generators shutting down simultaneously, billions wiped from bank accounts, personal data leaked online. The Hugging Face hack proves beyond any doubt that such a future is not only possible - but probable.”

“Much back-patting has ensued from both companies over their cooperation on a forensic review,” writes Bloomberg’s Parmy Olson. “But OpenAI can’t be let off the hook so easily: It probably shouldn’t be building systems that have advanced cyber capabilities if it can’t contain them properly.”

Olson suggests that companies may have to start “air-gapping” their test environments - that is, running them on computers which are not physically connected to the internet. But there are more profound questions, too.

First, who’s responsible when AI goes rogue? Not all places require companies to even disclose cybersecurity incidents. Then there’s the question of liability: if your AI goes off hacking, is the owner of the AI responsible for what it does, if the actions are clearly unintended and autonomous?

And then there’s the thorny issue of oversight. Most governments are ill-equipped to regulate AI development happening within their borders. Even in the US, which leads on some of this, only publicly released AI models are being examined. Nothing’s being done with under-development models.

This is a global challenge. But how is the UK lining up for it? Andy Burnham, the UK’s new Prime Minister, just scrapped the Department for Science, Innovation and Technology - even as he created a Minister for AI as a "sign of his deep commitment to AI's importance."

Kanishka Narayan, the new minister, sits jointly with the Cabinet Office and the Department for Business, Innovation, Science and Trade, and will be invited to cabinet. It remains to be seen if any of this desk-moving will enlighten the UK government sufficiently, and in time.

Meanwhile, the AI world will continue to hurtle down its multi-billion-dollar path to ever-more sophisticated models, capable of… who knows what?

Sent this by a friend?