BREAKING
DeepMind Unveils AI Control Roadmap
Layered Defenses Like Cybersecurity
1
Behavioral eval
↓
2
Monitoring & alerts
↓
3
Restrict / shutdown
↓
4
AI oversight model
0
M
Coding tasks analyzed
0
$M
Multi-agent safety fund
Attack Success Rates Quantified
Prompt injection
86
Sub-agent hijack
90
Data exfiltration
80
From Single Model to Multi-Agent
DeepMind
Layered
●
Staged safeguards
●
Supervisor AI oversight
●
$10M research fund
Anthropic
Zero trust
●
Treats agents as attackers
●
Industry-wide frameworks
Gemini Spark in Limited Beta
AI NEWS BLITZ
On June 18, DeepMind published a roadmap to defend against rogue AI agents.