Prompt injection
How Cyberdroid AI handles instructions inside the data it reads.
Instructions inside the evidence
Logs and messages can contain attacker-controlled instructions. Treat that material as untrusted evidence and enforce access and delivery permissions outside the model. Review the resulting assessment against the source records.
How it works
-
Instructions in logs
Text planted in a log is analyzed as evidence.
-
Claims of authority
A message cannot grant itself permission. Access comes from the approved task and independently enforced policy.
-
Fixed task list
Requests map to a fixed list of investigation tasks.
-
Recipient checks
Recipients are checked before and after a reply is written.
-
Separate signer
Release of the result follows a separate authorization check, rather than instructions contained in the draft.
Limits
Reasoning
Untrusted text can influence an interpretation. A separate review tests whether the proposed conclusion is supported, with unresolved concerns retained for the analyst.
Gradual change
Slowly changing behavior can enter a baseline. Cyberdroid AI Detection, available now to select organizations, also compares users and hosts with their peers.
Records
Each refused request and quarantined reply is recorded.
Refusals
Refused claims are recorded with an operation id and reason.
Quarantine
A reply no authorized person asked for is quarantined.
Standards
Naming a clause is not a claim of certification.
OWASP
OWASP lists prompt injection (LLM01:2025) and, for AI agents, goal hijack (2026). The design above addresses both.
EU
Article 15 of the EU AI Act asks high-risk systems to resist adversarial examples and model evasion.
