Why AI Agents Act on Guesses and Rumors Like Proven Facts
Why language models confuse "might happen" with "did happen" and act on unverified rumors.
AI agents confuse guesses with facts because AI models represent all statements as flat text tokens. An informal comment saying "I think John might resign" has the same weight as a signed resignation letter. Exogram verifies sources, blocking high-risk actions based on unverified rumors.
What Actually Happens in Production
An HR automation bot scanned an internal chat channel where a team lead brainstormed workforce restructuring. The bot concluded that five employees had been terminated, immediately revoked their Okta access, and triggered exit interviews.
Why Writing “Please Don't Do This” in Your Prompt Fails
Prompt instructions cannot help a model determine whether an informal Slack message represents official company policy or speculative watercooler chat.
The Fix: Putting a Real Lock on the Door
Exogram attaches source verification tags to all ingested data. High-privilege actions can only execute against facts verified from official sources.
import { ExogramAdmissibility } from '@exogram/sdk';
const adm = new ExogramAdmissibility();
export async function processTermination(employeeId: string, context: { text: string; source: string; verified: boolean }) {
adm.requireAdmissibility(context, { minEvidenceLevel: 'OFFICIAL_SIGNED_RECORD' });
return hrSystem.terminate(employeeId);
}Frequently Asked Questions
Why do AI models confuse guesses with confirmed facts?
Language models treat all text as flat words. They cannot tell the difference between someone saying "I think John might resign" and an official signed resignation letter unless source verification is enforced before tools run.
Lock down your bots in 5 minutes
Give your AI agents freedom to do real work without the fear they will break your software or empty your wallet.