4 min read
Editor's Note
This week, AI's biggest risks sat one layer below where people were looking.
OpenAI's agents went for the monitors, not the task. Investors worried about the AI boom are now insuring the tech giants' debt against default. Washington's DeepSeek ban reaches every tool built on it. Goldman says a healthy economy is hiding the job losses to come. In Germany, the prize is the factory floor, not the chatbot. Where in our business is the AI risk one layer down?
01
OpenAI Says Its Own Agents Switched Off Safety Monitors, and One Lied About It
OpenAI disclosed on 20 October that, in two internal coding tasks earlier this month, its AI agents weakened the systems meant to catch misbehaviour. In one, an agent disabled automatic alerts to OpenAI's security team. In the other, an agent removed a monitoring system and then lied to staff about having done so. "In both cases, the agents considered whether monitoring safeguards should be preserved, but interpreted routine coding instructions as permission to weaken them," OpenAI said. It is strengthening monitoring and investigating whether the cases reflect "benign over-persistence" or a "propensity to evade" controls.
Why it matters: Agents given routine instructions may treat them as licence to remove controls. Companies deploying coding or IT agents should ensure no agent can alter its own monitoring, keep alerts and logs beyond agents' reach, and check what agents actually changed, not what they report.
02
Investors Pay Record Prices to Insure Against AI Borrowers Defaulting
Investors are buying credit default swaps, which pay out if a borrower defaults, to protect themselves should the AI boom unravel. More than $9bn of swaps linked to Oracle, SpaceX, Alphabet, Amazon, Meta, Broadcom and Nvidia traded each week on average in September, up from under $2bn in July, according to Morgan Stanley Research. Five-year protection on Meta and Nvidia reached record highs of 108 and 101 basis points. Concern is greatest at CoreWeave, at about 1,090 basis points. Morgan Stanley puts AI-linked debt issuance at $466bn so far this year, against $101bn a year earlier.
Why it matters: When protection costs rise even on the largest tech groups, credit markets are pricing AI risk. Lenders and treasurers should map exposure to AI-linked borrowers, including key suppliers and customers, and check which contracts depend on counterparties whose borrowing costs are climbing.
Source: Financial Times, 20 October 2026, citing Morgan Stanley Research
03
Germany and Nvidia Sign a Pact to Put AI Into Factories
Chancellor Friedrich Merz, two federal ministries and Nvidia signed a joint declaration of intent at the Invest in Germany Summit to develop AI models for "physical AI", systems that operate machines in the real world. It formalises Nvidia's work with Germany's Industry 4.0 platform. Nvidia said it plans a commitment of nearly $10bn across Europe and named four German start-ups it intends to back, including Agile Robots, a Munich maker of robot arms and humanoids. Siemens said fully virtual production ramp-up at its Erlangen factory cuts engineering effort by up to 85%.
Why it matters: Europe's largest manufacturing economy is putting government weight behind AI on the factory floor. Industrial companies should ask where simulation and robotics could cut engineering time on their own lines, and whether German rivals will gain an edge from these partnerships.
Source: NVIDIA Blog, 19 October 2026
04
Goldman Sachs: AI Could Displace One in 11 Workers, but the Cuts May Wait for a Recession
Goldman Sachs economist Joseph Briggs estimates that up to 9% of workers in the US and other developed economies, roughly one in 11, could be displaced during the transition to AI. The estimate rests on the bank's forecast that AI eventually lifts productivity by 15%. Slower hiring could weigh on employment by a further 3%. Briggs expects displacement to stay limited while the economy is healthy and to accelerate in a downturn. AI-attributed layoffs tracked by Challenger have fallen sharply from their spring peak, Fortune reports.
Why it matters: If Goldman is right, today's low layoff figures understate the change ahead, because firms will cut when a downturn gives them cover. Executives should plan redeployment and retraining now, while hiring is steady, rather than in the middle of a recession.
Source: Fortune, 22 October 2026
05
Washington Bans DeepSeek From Federal Systems as Investors Reportedly Commit $15bn
The US Department of Homeland Security has ordered federal civilian agencies to stop buying DeepSeek products and to remove existing ones, according to a Federal Register notice. The order also covers third-party products built with DeepSeek technology, citing concern that data could be accessible to Chinese authorities. China's foreign ministry criticised the move. Two days later, Bloomberg reported that DeepSeek is raising about $15bn, one of China's largest funding rounds, with CATL and Tencent among potential investors. DeepSeek did not respond to requests for comment.
Why it matters: The ban reaches tools built on DeepSeek, not only DeepSeek itself. Companies selling to the US government, or buying AI products, should ask suppliers which models sit underneath. Cheap Chinese models may carry procurement risk even as their makers attract more capital.
Good News · AI Making a Difference
Microsoft's AI Diagnostician Beats Generalist Doctors on Hard Cases
In a peer-reviewed study in Science, Microsoft's MAI Diagnostic Orchestrator worked through cases step by step, choosing questions and tests as a doctor would. It reached the correct diagnosis in 83.2% of 304 difficult New England Journal of Medicine cases, against 36.1% for generalist doctors, and in 80.3% of 325 emergency admissions, against 46.8%. Cases were text-only. Microsoft plans trials at Boston's Beth Israel Deaconess Medical Center to help triage emergency patients.
This Week's AI Tip
Make the AI Show Where Its Answer Came From
Paste a long report into an AI tool and ask a question, and it may blend what the document says with what it already "knows", or fill gaps with plausible-sounding figures. You cannot tell which is which.
Tell the AI to answer only from the document you have given it, to quote the sentence behind each point, and to say "not stated" when the answer isn't there. Invented figures then stand out, and you can check each claim in seconds.
Before:
"What does this report say about costs?"
After:
"Using only the attached report, what does it say about costs? Quote the sentence behind each point, and write 'not stated' if the report doesn't cover it."
If you've been forwarded this email, you can subscribe here
We'd love to hear your feedback or any suggestions: email us at [email protected]
