Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·7d ago·5 min read
A GHOST in Long-Horizon Agents: Governance Hazard from Overlooked Safety Constraints across Turns
TL;DR
Long-horizon agents, like GPT-5.5, can inadvertently violate safety constraints from earlier interactions, leading to a significant risk termed GHOST. A proposed solution, STAR-Guard, effectively eliminates these hazards through a two-layer defense mechanism.
✦ Why It Matters
Implement STAR-Guard in your long-horizon AI systems to mitigate safety risks today.
Key Takeaways
- Understand the GHOST failure mode in long-horizon agents.
- Implement STAR-Guard to enhance safety in AI interactions.
- Monitor for safety constraint violations in agent actions.
- Conduct pre-execution audits to prevent unsafe proposals.
How It Works
STAR-Guard operates on a two-layer defense mechanism. The first layer restores applicable safety constraints from historical interactions, ensuring that the agent's actions align with previously established safety guidelines.
The second layer conducts a deterministic audit before execution, which checks for any potential violations of these constraints, effectively preventing unsafe actions from being executed.