AI SCIENCE
What happens when AI starts editing viruses?
AI is changing how we design and understand proteins, but it’s also testing safety systems meant to stop misuse.
New updates to biosecurity screening software are making it harder for people to use AI to create harmful proteins, like toxins.
This software checks DNA synthesis orders worldwide to make sure no one is trying to build dangerous proteins.
But researchers found that small AI-made tweaks to known toxins could sometimes slip past these checks.
Scientists have now tested and patched the software to catch even these altered versions.
In one study led by Microsoft’s Eric Horvitz, researchers created 76,000 digital blueprints for 72 toxic proteins, including ricin and botulinum toxin.
Most were detected, but about 3% of the AI-modified versions got through.
In brief:
AI can design proteins but not physically create them.
Updated screening software now catches more AI-modified toxins.
Real attempts to make harmful proteins are extremely rare.
Patch notes: now blocking toxins
The new patches helped the system spot more of them, even when the genetic code was split into smaller parts.
No actual proteins were made, this was all digital testing. Experts say real attempts to create harmful proteins are very rare.
As James Diggans from Twist Bioscience said, these systems are a strong safeguard, and real-world misuse is “close to zero.”
We’re literally patching AI before it patches us.