Hello, and welcome to the latest AI Update. A special note for this edition: We have just launched a new site focused on AI security, along with a dedicated newsletter – Control Plane. If you enjoy ID Tech’s weekly AI Update, please go ahead and sign up for the newsletter, where I’ll be providing even more in-depth coverage of the AI security space.
Now, on to this week’s Update:
Anthropic is preparing to launch a powerful new AI model that has cybersecurity implications. In an apparent blog post available through publicly available Anthropic assets, which has since been removed, the company explained that the new “Mythos” model “gets dramatically higher scores on tests of software coding, academic reasoning, and cybersecurity, among others,” compared to its currently available flagship model, Opus 4.6. The post went on to explain that the model has prompted Anthropic to take a more careful approach to safety before proceeding with a large-scale release. “We’ve documented the ways in which models can be used to rapidly discover vulnerabilities in codebases; we’ve also shown how they’re already being used to commit large-scale cyberattacks,” the post said.
Meanwhile, the firm has succeeded in getting an injunction to block the Pentagon’s enforcement of its “supply chain risk” designation for Anthropic over a dispute about how its technology could be used by the Department of War. A federal judge in San Francisco ruled that the designation was “likely both contrary to law and arbitrary and capricious.” The ruling doesn’t mean that the DoW must continue contracting with Anthropic: the government is free to choose not to do so. It just can’t use this blacklisting mechanism in this way. The order is stayed for seven days to give the Pentagon time to appeal.
China’s largest chipmaker has allegedly been supplying chipmaking tools to Iran for about a year, according to a new Reuters report, which cites unnamed US officials. SMIC’s semiconductor technology is suspected to have been provided to Iran’s military, along with some training. SMIC has been on the U.S. Entity List since December of 2020, with the Commerce Department arguing at the time that it was prudent to limit SMIC’s access to American tech given its ties with the Chinese military. It may be an information op, given the thin sourcing, but it’s nevertheless noteworthy that American officials are floating the allegation amid chatter about a potential escalation of the current conflict in Iran.
Domestically, China appears to be putting the brakes on Meta’s deal to acquire the AI agent startup Manus, launching an official review of the agreement and temporarily barring Manus’ co-founders from leaving the country. Chinese authorities reportedly questioned the executives over alleged violations of foreign direct investment reporting rules. More broadly, they may be concerned about a potential brain drain as American tech companies seek to acquire Chinese AI tech and talent. This may also be part of an emerging phenomenon in which government authorities seek greater control over AI technology that is increasingly seen as tied to state power and even national security.
OpenAI is shutting down its Sora app, which briefly captivated users through its highly realistic video output and high-fidelity deepfakes of users who opted in with their faces. The problem was that this output was very compute-heavy, and Sora wasn’t making much money at a time when OpenAI needs to tighten up its focus on profit while aiming for an IPO. There have also been scattered reports of Sora’s popularity as a vehicle for disinformation, though it’s not clear whether this factored into the decision to shut it down. In any case, the move surprised a lot of people – including some OpenAI employees as well as Disney, which had been working with OpenAI on a $1 billion licensing deal for the use of its characters on the platform.
Canada’s Liberal government has proposed legislation aimed at the use of deepfakes to affect election results. Among other things, the Strong and Free Elections Act would make it illegal to knowingly spread false or misleading information about election activities or the voting process with the intent to disrupt an election or affect the outcome, and on that note, it specifically targets sophisticated likenesses, though it never actually uses the word “deepfake”. Rather, it says if a person “creates or causes to be created an image that is likely to be mistaken for an image of any of those persons” with the intent of influencing the election, that’s a crime. There is similar language about voices. It will next face a second reading in Canada’s Parliament before further study on the road to becoming law.
Finally, have you ever seen one of these?
It’s called a Flipper Zero, and it’s basically a Swiss Army knife for hacking digital signals to do things like copy key fobs and hijack public television screens. And an expert LLM “jailbreaker” has now created software that allows you to interact with the device like you would with a chatbot. I wrote about this in the latest Control Plane newsletter, which you can check out here.
Thanks for reading!
–
By Alex Perala
