Researchers at Microsoft have used artificial intelligence to find a previously unknown weakness in biosecurity systems, MIT Technology Review reports (unpaywalled copy). These systems screen DNA orders to block the creation of dangerous toxins or pathogens.
A zero-day vulnerability is a hidden flaw that defenders do not know about, like an unseen hole in a fence. This work is published in Science.
Generative artificial intelligence (AI) helps find new medicines but can also make harmful molecules. The Microsoft researchers targeted biosecurity screening software used by DNA vendors. To make a protein, scientists order DNA sequences from companies. The software checks orders against lists of known threats and alerts if there is a match. Using AI models like their own EvoDiff, the Microsoft researchers redesigned toxins, changing the DNA structure just enough to evade detection but keeping the toxin's deadly function. The tests were all digital, with no real toxins made, to avoid any hint of weapon development.
Patches issued but challenges remain
Before sharing results, Microsoft warned the US government and software companies. These groups created patches to fix the flaw, though some AI-made molecules still slip through. This could become an ongoing struggle, like an arms race between attackers and defenders. To prevent misuse, the researchers hid some code and did not name specific toxins, though examples include ricin from castor beans and prions that cause mad-cow disease.
This underlines the need for better DNA screening with strong checks. The US government sees this as key security, with calls for updates, though new rules are pending. Some experts disagree on fixes. Disguises will always exist, some experts say, and tests should be tougher. Perhaps safety should be built into AI itself, like limits on what it shares, since AI training is hard to control widely. Overall, the discovery shows AI's power in biology demands constant vigilance for biosecurity.