Email: [email protected]
Researchers have demonstrated how malicious payloads can be hidden within just 0.00001% of an AI model's weights, creating 'sleeper agents' that bypass traditional security inspections and trigger under specific conditions.