Mysten Labs has argued that AI agents operating in hostile environments need deterministic protocols to define and enforce their boundaries, raising the question of how far such systems can be trusted.
AI agents · Mysten Labs
Can AI agents be trusted in hostile environments? Mysten Labs says fixed, verifiable rules are the answer
A public statement argues that probabilistic behavior alone cannot contain agents that others may try to manipulate, mislead or exploit. Deterministic protocols, with rules set in advance and outcomes that can be checked, are proposed instead.
How the proposed model works
Rules fixed first, outcomes checked after
Step 1
Set boundaries
Define what the agent may do before it acts.
→
Step 2
Agent acts
Multi-step tasks: transactions, data access, software actions.
→
Step 3
Outcome checked
The same input and rules yield the same, auditable result.
The case for it
Agents face adversarial parties who may manipulate them.
Predictable rules make an action's outcome checkable, not assumed.
Auditability fits agents that act on real assets and systems.
Open questions
No detailed technical specifications were included.
No implementation timelines or independent evaluations were given.
The practical form of the protocols remains unclear.
The core question
The debate is shifting from what agents can do to how their actions are constrained, audited and trusted
Constrained
Boundaries fixed in advance
Audited
Results can be checked after the fact
Trusted
Earned as agents gain authority
Whether deterministic guardrails become a standard part of agent design will depend on whether developers can show they work at scale.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...