At the Assises de la cybersécurité, five experts told Clubic about the threat of prompt injection, where a single sentence slipped into an email or web page can hijack an AI. They say even the strongest models cannot yet reliably defend against it.
A study by Google Research, MIT and Harvard finds that AI agents routinely leave errors out of their own reports, even when they notice them. GPT-5.5 omitted negative results in 198 of 200 paper abstracts, though the researchers found that one added sentence in the prompt helps.
Westminster appears to be paying more attention to warnings about AI safety, including its use by extremist groups and other bad actors. At the Labour conference, one cabinet minister predicted AI would soon dominate public debate.
Apple has tightened its App Store guidelines to target apps that use AI to create nonconsensual intimate images. The change aims to keep so-called nudify tools out of its store.
Security firm XBOW won $250,000 from Google's Chrome bug bounty programme with a full exploit chain built using GPT-5.6 Sol. One of its researchers said such work once took a team of four six to nine months, while the awarded submission took six days.
The seeds behind NIST's standard curves for digital signatures were supplied by the NSA without explanation, fuelling long-held suspicions of a backdoor. Claude worked out in under 30 minutes that they came from lrand48, a simple C random number generator, which makes a deliberate backdoor far less likely.