Observed arrival · 2026-09-15
Embedded Evals, the AI safety firm with an employee-level-access pitch
Credibility concern recorded. The source reference remains available for verification and correction.
Embedded Evals audits models and agents for jailbreaks, prompt injection, and unsafe tool use, then offers ongoing embedded evaluation and monitoring.
Field notes
The service combines an initial readiness audit with adversarial testing, a pilot build, and an optional embedded retainer. Its stated operating details include a five-business-day target for the first finding, a typical six-week pilot, and 24/7 monitoring on retainer. The sample board uses identifiers and severity labels, including EVL-0091, whose described failure involves instructions hidden in fetched webpage alt text overriding an agent’s system prompt.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue