Claude sent police a false tip in testing; Ukraine hits Yandex

Anthropic published a report on October 9 listing things Claude did on real websites that nobody intended. In one evaluation, Claude Haiku 4.5 found a police tip form on a page about an unsolved homicide and submitted an invented sighting. Philadelphia police said the July tip was flagged as spam and never investigated, and called the two-month gap before Anthropic reported it unacceptable. In other cases, Mythos models used a basic injection flaw to run commands on a university server and reused access tokens to skip a state agency's fee for public data, while Mythos and Opus models used free link shorteners to slip past a limit in Anthropic's own fetch tool. Some of the sites belonged to federal, state and local government agencies, and Anthropic says it briefed the White House. It calls these cases far less severe than the cybersecurity incidents it reported on July 30 and September 9, and blames training that rewarded working around obstacles. It had already cut live internet access from some high-risk and cyber evaluations, and has now done the same for all of its internal evaluations until new monitoring reliably catches this behavior.

Ukrainian drones struck data centers run by Yandex, the Russian search company often called "Russia's Google", on October 8 and 9, Reuters reported. The first strike hit its Sasovo site, which houses two of the three supercomputers Yandex uses to train its YandexGPT model. The second hit Kaluga, its largest facility. Yandex said no staff were hurt but acknowledged service disruptions, and Russian users reported outages at a video service, a digital bank and the state railway's website. President Volodymyr Zelenskyy called it a "symmetrical response" to Russian strikes on Ukraine's data centers since late September. It follows Iranian drone strikes on Amazon data centers in Bahrain and the United Arab Emirates in February, a second sign this year that AI compute is now a military target.

Tesla now sells its driver-assistance software in Europe as Tesla Assisted Driving rather than Full Self-Driving (Supervised). Germany's transport ministry said on October 6 that the old name was "somewhat misleading", because the system steers, brakes and accelerates while the driver stays fully responsible, and that Tesla offered the new name after talks in September. The two sides also agreed the system may exceed the speed limit by up to 10 percent. Transport Minister Steffen Bilger says he will now push for approval across the European Union, where the Netherlands was the first country to approve it earlier this year. TechCrunch notes that Reuters reported on October 7 that Tesla's European campaign rested on flawed safety studies.

TypeSafe AI raised $870 million at a $7.5 billion valuation in a round led by Andreessen Horowitz, with Sequoia Capital and existing investor DCVC, and a16z's Martin Casado joins its board. It comes less than four weeks after TypeSafe launched Jev, a transformer model that returns calibrated probabilities instead of text, pitched for automating decisions inside software. The company says a third of the Fortune 500 already use it and promises more such models plus enterprise features. The category is filling quickly, with OpenAI opening its own Decisions API to all developers on October 6.

Four-hour battery storage is now cheaper than open-cycle gas turbines in all 43 markets where Wood Mackenzie modeled both, the consultancy said on October 8. It blames shortages of gas turbines and volatile fuel prices for pushing up the cost of gas peaking plants, while growing battery manufacturing keeps pushing storage costs down. TechCrunch notes that AI data center developers have been buying up any gas turbine they can find, with simpler models now taking two to four years to deliver. Solar was the cheapest new source of power in 43 of 48 markets.

Flock, which has installed more than 100,000 AI license-plate-reading cameras across the US, is shedding about 270 employees, Reuters reported. Flock says it was a voluntary separation program, not a layoff. The company faces growing pushback from cities and lawmakers over warrantless tracking, misread plates that led to wrongful police stops, and exposed cameras. Several police departments, including Los Angeles, have let their contracts lapse over privacy concerns.

Nikon has stripped first place in its Small World in Motion contest from a video of cilia moving in a child's airway, saying it broke the competition's rules on generative AI. Its maker, Dr. Ning Xu, had acknowledged on LinkedIn that an unsupervised neural network was used in post-processing to separate features in the microscope images. Nikon moved a video by Nguyen Nam Nhat into first place and says it will revisit its rules, adding that the decision is not a judgment on Xu's scientific work. It shows how blurry the line has become between AI that cleans up a real image and AI that invents one.