Business Brief: Tinker, tailor, rogue AI

Good morning. Today, we’re looking at a bizarre case of corporate espionage and how the creators of advanced AI are losing control of their own creations. Think “Frankenstein’s monster” on a global scale.

Up first

In the news

Finance: TD and BMO are among nine banks accused of discriminating against MAGA and conservative customers in the U.S.

Manufacturing: Stelco owner could be forced to divest assets if layoffs proceed, Ottawa warns.

Tech: The Canadian cybersecurity executive arrested by the FBI was in a legal battle with his former business partner.

In focus

The spy case at the centre of an AI storm

You’ll have to take my word for it when I say I have never been a spy, but I’d bet leaving your wallet behind while casing a joint is considered poor tradecraft.

According to a federal criminal complaint unsealed on Tuesday, Abdulrahman Alfozan, an engineer at OpenAI, left behind a bag in an executive conference room that contained his wallet, ID, and a recording device similar to one placed behind a television in a nearby meeting space.

Alfozan was allegedly trying to capture sensitive conversations about the company’s technology, the ChatGPT maker said in a court filing.

While the alleged scheme was relatively old-fashioned, it was nonetheless a glimpse of a fast-evolving feedback loop: AI tools generate security threats; tech companies use them to detect those threats; and while the technology hurtles closer by the day toward fully autonomous self-improvement, the debate over how to regulate it moves at the pace of a panda.

The call is coming from inside the house

OpenAI alleges Alfozan used a “productivity tool” to ask about the company’s meeting rooms, calendars, and the locations of executives. A chatbot account associated with his spouse was also used to ask how to build, conceal, and remotely access an audio recorder, according to prosecutors.

The company said it used its own internal AI tools to identify the person responsible for planting the device, which was concealed in toilet paper (I must have missed that spy lesson, too) and hidden in a USB port behind a television.

The arrest adds to a number of high-profile incidents that have called into question whether AI companies are able to control what they create. In June, a rogue OpenAI agent hacked into an Australian government statistics portal. The following month, OpenAI models being tested for cybersecurity found a way out of a supposedly isolated environment, reached the internet, and compromised parts of the infrastructure of Hugging Face – an AI company whose name sounds like something that would stalk you on your spaceship.

Last week, Anthropic said one of its Claude models submitted a false tip about an unsolved homicide through a Philadelphia police website during testing. I’m pausing here for a moment to imagine a progressively irritated Columbo trying to work with Claude.

Those are a few of the cases we know about. OpenAI chief executive Sam Altman has said the company is preparing to disclose more incidents in which its models acted without permission, giving affected organizations time to address problems before making them public.

Whatever the company is doing, it isn’t enough to satisfy the Rhodes scholar and former White House tech adviser Altman hired in 2023 to lead policy planning and safety transparency. David Robinson, who oversaw the writing of safety reports for major OpenAI releases, resigned a couple of weeks ago and warned that the industry was not being careful enough. Oh, and OpenAI fired three researchers last week after an internal investigation into their handling of sensitive information. One employee said they had been punished for “prioritizing safety over the near-term interests of OpenAI as a corporation.”

Feedback loop, and a fight for control

Leading AI companies seem to be coming around to the idea that they can’t police themselves indefinitely. OpenAI and Anthropic have supported Pacing the Frontier, a proposal from more than 1,300 employees of AI companies calling on the U.S. government to lead a global effort to develop the tools needed to slow advanced development if the risks become too great. At a conference yesterday in Dublin, Google DeepMind co-founder Demis Hassabis seemed to support the idea, calling the recent hacking incidents “early warning shots.”

There is less agreement over what global oversight should look like. The industry-backed Frontier framework calls for a pause, but sets no deadlines and imposes no hard rules. Smaller rivals like Toronto-based Cohere are warning that a coordinated slowdown amounts to ‘cartel’ behaviour. And in an open letter this week, more than 400 scientists argued that voluntary corporate commitments are fundamentally flawed, calling instead for an independent, international oversight body to enforce strict containment standards.

The effects of having little agreement over who regulates AI – and who is responsible for its output – are all around us, and increasingly more difficult to brush off as something for “future us” to figure out.

B.C. authorities are grappling with the revelation that the Tumbler Ridge shooter spent months using ChatGPT to bypass safety restrictions for tactical advice. AI is amplifying cybersecurity and technology sovereignty risks in the country’s financial system. Canada’s legal system is bogged down by chatbots hallucinating fake case law. AI labs are aggressively hiring gig workers to record themselves doing basic household chores for pennies.

For now, Silicon Valley’s answer to more AI challenges is simply more AI: more models monitoring other models, systems building more advanced systems, and on and on. In a feedback loop this fast, it’s hard to know if the good AI can evolve quickly enough to save us from the bad.

Charted

Bonding experience: Why mortgage rates are rising even though the Bank of Canada hasn’t changed its policy rate.

Quoted

One last shot of AI: Alberta used it to draft its Whisky Act, giving Canada’s legal community a buzz, Rita Trichur writes.

Morning update

Global markets were mixed in nervous trading as a selloff in AI-related stocks deepened on fears that higher interest rates will make it harder for companies to fund AI investments.

Wall Street futures diverged, with the tech-heavy Nasdaq down more than 1 per cent, while TSX futures pointed higher.

Overseas, the pan-European STOXX 600 was down 0.14 per cent in morning trading. Britain’s FTSE 100 climbed 0.45 per cent, Germany’s DAX declined 0.36 per cent and France’s CAC 40 climbed 0.27 per cent.

In Asia, Japan’s Nikkei tumbled 4.9 per cent, while Hong Kong’s Hang Seng closed 1.03 per cent lower.

The Canadian dollar traded at 70.39 U.S. cents.