Model poisoning embeds hidden sleeper agents into AI weights. Microsoft has launched a detector that identifies three red flags: unnatural attention on specific words, a memorization bias toward malicious data, and fragmented triggers that activate backdoors despite typos or partial phrases.
Read more details: https://rebrand.ly/model-ec6700
Comments