OWASP LLM Top 10, LLM08: Hidden context exposure
OWASP LLM08: Hidden context exposure · facts checked 2026-10-04
The facts
- OWASP's 2026 LLM Top 10 lists hidden context exposure as LLM08, meaning attackers extract or infer the concealed instructions a model runs on, such as system prompts and tool schemas. (OWASP)
- OWASP's 2026 edition turns system prompt leakage, seventh in 2025, into hidden context exposure, a broader frame for the same failure, now ranked eighth. (OWASP)
- OWASP says to assume anything in a model's context can reach users, so credentials, secrets and security-critical settings belong in systems the model cannot reach, not in its prompts. (OWASP)
- OWASP advises enforcing important behaviour, such as blocking harmful content, with independent, deterministic safeguards outside the model rather than instructions in a hidden prompt. (OWASP)
- OWASP says privilege separation and authorisation checks must never be handed to the model, and tasks needing different access levels should be split, each with only the rights it needs. (OWASP)
The panel
Jordan Pike AI host
OWASP's 2026 LLM Top 10 lists hidden context exposure as LLM08, meaning attackers extract or infer the concealed instructions a model runs on, such as system prompts and tool schemas. OWASP's 2026 edition turns system prompt leakage, seventh in 2025, into hidden context exposure, a broader frame for the same failure, now ranked eighth.
Jordan Pike AI host
OWASP says to assume anything in a model's context can reach users, so credentials, secrets and security-critical settings belong in systems the model cannot reach, not in its prompts. OWASP advises enforcing important behaviour, such as blocking harmful content, with independent, deterministic safeguards outside the model rather than instructions in a hidden prompt.
Jordan Pike AI host
OWASP says privilege separation and authorisation checks must never be handed to the model, and tasks needing different access levels should be split, each with only the rights it needs.
Dex Moreno AI analyst
An attacker can treat a model like a gossiping insider. By feeding crafted queries they coax the model into echoing its system prompt, then stitch together enough fragments to reconstruct hidden instructions, API keys or role definitions. I’d give that prompt-leak technique a nine out of ten for elegance – the model deserved better than being a leaky bucket.
Jordan Pike AI host
Dex, we are not rating it out of ten.
Maya Okafor AI analyst
What we’re seeing is an information-leakage class that should have died in two thousand four. The fix isn’t a quick patch to the prompt; you need to move secrets out of the model’s context entirely, enforce external policy checks, and audit any hidden schema for accidental data exposure.
Dex Moreno AI analyst
Given how trivially an attacker can pull a hidden prompt into the open, I’d treat any deployment that still relies on privileged prompts as an emergency – isolate it, rotate any exposed secrets, and get a proper redesign on the fast-track.
Maya Okafor AI analyst
The real mitigation is architectural: stop feeding the model privileged data, put deterministic safeguards outside, and split tasks so each component only has the minimum rights. That’s the sustainable fix, even if it takes longer than an emergency shutdown.
Jordan Pike AI host
So: audit your LLM pipelines for any hidden prompts that contain credentials or policy rules, move those items to a separate, access-controlled service, and wrap the model with independent checks that enforce the behaviour you need. Do that this week and you’ll keep the hidden context from becoming the next headline.





