Access and authorization
Authentication, scoped permissions, session handling, and tool-level authorization reduce the chance that an agent or compromised account can reach more than it needs.
LegionASI / Safety
Useful systems need clear boundaries.
Safety is an engineering discipline around the model, the interface, the data, the tools, and the people who authorize action. LegionASI language models can be wrong, overconfident, manipulated, or incomplete. A safe design makes those limitations visible and limits what a mistake can do.
Safety principles
Higher autonomy requires stronger controls.
An agent that drafts text has a different risk profile from an agent that can change records, send messages, spend money, or control an external system. We treat autonomy as a permission that must be earned, scoped, monitored, and revocable.
Human control
People remain responsible for consequential decisions. The system should make review and correction easy rather than presenting an automated recommendation as a final authority.
Permission boundaries
Agents should operate within the user’s stated task and the permissions granted to that task. They must not silently expand their objective, access, or audience.
Explicit authorization
External actions require explicit, scoped authorization. Drafting an email is different from sending it; proposing a change is different from applying it.
Accountability
Important actions should be attributable to the responsible user, agent, tool, and decision point so that failures can be investigated and corrected.
Engineering controls
Controls are designed around failure modes.
Abuse prevention
Usage limits, request validation, policy enforcement, and monitoring help address spam, malicious automation, prompt abuse, and attempts to bypass safeguards.
Data handling
Conversation context and uploaded material should be limited to the task, handled according to the service’s privacy practices, and not exposed to an agent that has no reason to see it.
Auditability
Logs and review records can help reconstruct what was requested, what the system produced, what tools were invoked, and where a human approved an action.
Rate limiting
Limits protect service availability and reduce the blast radius of runaway loops, accidental repetition, credential misuse, and excessive automated requests.
Security testing
Testing should include authentication paths, input handling, prompt injection, unauthorized tool use, data isolation, model escalation, and failure recovery—not only normal chat quality.
Model limitations
Confidence is not correctness.
Language models can produce a fluent answer that contains a false statement, misses important context, follows the wrong interpretation of a request, or repeats an error introduced earlier in a multi-agent chain. A confidence signal is not proof. Factual claims that matter should be grounded in suitable sources and independently checked.
For sensitive domains and consequential work, users should treat AI output as assistance for review—not as a substitute for qualified professional judgment. When an agent is uncertain, lacks evidence, or encounters conflicting instructions, the safer behavior is to surface the conflict, ask for clarification, or stop.
Safety is ongoing. We expect controls, evaluations, and product behavior to change as LegionASI adds capabilities. If you identify a security or safety concern, use the Contact page and describe the affected behavior without sharing secrets.