Microsoft Finds Way to Hide Backdoors in AI Helpers
This could make free AI tools riskier to use—here's what you need to know.
Be careful if you’re running sketchy quants/finetunes. If these techniques become more widely repeatable, “sleeper agents” could genuinely develop into an attack vector. What are your thoughts?
- Microsoft showed they can backdoor an AI by changing fewer than 10 internal settings, making it act maliciously later.
- This means free AI models from untrusted sources could be secretly compromised and turn on you.
- Stick to official AI providers and avoid downloading sketchy AI software to protect yourself.
Why It Matters
Using AI from untrusted sources could put your data and devices at risk—stick to reputable providers.