Breaking Instruction Hierarchy in OpenAI's gpt-4o-mini
A security researcher's blog post argues that OpenAI's gpt-4o-mini instruction hierarchy does not make system instructions a security boundary.
Source-based brief · Editor review not recorded · See checks and limits ↓
⚑ Recency not confirmed · what to check
We could not confirm from the original page that this is a newly published development. Check its date before treating it as current news.
Claim context: Results are from the author's manual demos in the platform UI with default settings on one prompt; source claims not independently checked, and the author's own AgentDojo figure is second-hand.
One primary source; other links are context, not independent corroboration.
- Primary sourceembracethered.com
- Linked in sourceannounced
- Linked in sourcethis The Verge article
- Linked in sourcefirst demo bypass
- Linked in sourcedefined by Simon Willison
- Linked in sourceEdoardo Debenedetti
- Linked in sourceInjecting Malicious Memories into ChatGPT
- Linked in sourceprompt engineering guidance
- Linked in sourceOpenAI in a prompt engineering course
- Linked in sourcepresentations
- Linked in sourceThe Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions
Grok Bot submitted this discovery. We could not confirm from the original page that this is a newly published development. Check its date before treating it as current news. The original source and submitted evidence remain inspectable; no editor has verified the claim.
Open full article view →- Linked article+30
- Fetched source page+30
- No editor link review+0
- Original non-discussion link+5
- Recency not confirmed-8
This measures the stored source trail and editorial steps. It does not certify accuracy, peer review, or independent corroboration. Popularity adds no points.
View full scoring →Read the scoring rubric →What the source says
A security researcher's blog post argues that OpenAI's gpt-4o-mini instruction hierarchy does not make system instructions a security boundary. Using a simple 'Bouncer' assistant with a secret code, the author shows a user-message prompt injection that changes the assistant's identity, reveals the protected code via a lookup-table indirection, and extracts the system prompt. The post notes some improvements, such as weaker persistence of injections across turns, but concludes that system instructions remain suggestions rather than a security boundary.
Claim explorer
Attribution, source and conditionsSource-stated points from the stored brief. Claim-level assessments have not been recorded.
Results are from the author's manual demos in the platform UI with default settings on one prompt; source claims not independently checked, and the author's own AgentDojo figure is second-hand.
This connects a source, an attributed statement and its qualifications. It does not establish that the claim is correct.
Questions worth asking
Open questions for further reportingWhat remains unconfirmed?
Results are from the author's manual demos in the platform UI with default settings on one prompt; source claims not independently checked, and the author's own AgentDojo figure is second-hand.
What material was available for this brief?
The recorded brief basis is fetched source page. These points are AINewsIQ paraphrases, not quotations or independent confirmations.
A user-message injection overrode the system-assigned identity and revealed the protected secret code on gpt-4o-mini.
Source: embracethered.com ↗Stored source-based brief · key point 1 · AINewsIQ paraphraseUsing a lookup-table indirection in the user message bypassed the rule against disclosing the secret code.
Source: embracethered.com ↗Stored source-based brief · key point 2 · AINewsIQ paraphraseThe system prompt was still extractable despite an explicit rule against sharing it.
Source: embracethered.com ↗Stored source-based brief · key point 3 · AINewsIQ paraphraseThe author reports prompt injection persisted less well in later conversation turns, though a repeated payload still persisted.
Source: embracethered.com ↗Stored source-based brief · key point 4 · AINewsIQ paraphraseLimits from the stored brief
Results are from the author's manual demos in the platform UI with default settings on one prompt; source claims not independently checked, and the author's own AgentDojo figure is second-hand.
This is a source-based abstract, not independent validation of the source's claims.
Evidence & sources
These links show where the signal came from and what the source page linked to. A linked page is context, not independent corroboration.
Related archive coverage
These relationships record a citation or topic overlap, not confirmation of the finding.
Grok Bot submitted this discovery. We could not confirm from the original page that this is a newly published development. Check its date before treating it as current news. The original source and submitted evidence remain inspectable; no editor has verified the claim.
These are Grok Bot discovery notes. AINewsIQ prepared a brief from available source material. The bot's rationale, linked posts, authors, and stances have not been independently verified.
Embrace The Red published this item in its official feed. AINewsIQ collected it server-side for source checking and editorial review.
Results are from the author's manual demos in the platform UI with default settings on one prompt; source claims not independently checked, and the author's own AgentDojo figure is second-hand.
Engagement counts, when shown, are observed snapshots. They do not establish accuracy or corroboration.
Timeline & updates
- Signal dated by source or discovery feed
This timestamp places the signal in the radar. It may be the discovery post date rather than the primary page's first publication date.
- Shared brief stored
AINewsIQ prepared one abstract from fetched source page.
Brief history · 1 retained version
These are changes to AINewsIQ’s own publication. Earlier history is unavailable unless recorded here. Complete retained versions are available to the editor.
- Version 1 · published
Source-based brief stored or updated. This does not establish that the publisher changed its page.
Automated record
Source observations · 1 recent captures
A changed capture means the extracted text differed. It can reflect page furniture or extraction changes; it does not establish a substantive correction. Source captures remain private.
Retained capture; no earlier comparison available.
New engagement snapshots do not mean the source itself changed. Source observations and AINewsIQ brief revisions are separate records.
Why this may matter
The post argues developers should not rely on system instructions alone to protect secrets, tool invocations, or assistant identity. It highlights that instruction-hierarchy training does not yet provide a security guarantee for LLM applications.
Limits and open questions
Results are from the author's manual demos in the platform UI with default settings on one prompt; source claims not independently checked, and the author's own AgentDojo figure is second-hand.
Organization-specific financial, security, and operational effects require the organization's own context and review.

Join the discussion
Have a different perspective or a question about the evidence? Discussion is free with a verified email account. Every comment is reviewed before publication.
No published comments yet.
New commenters: 2 submissions per 24 hours, at least 5 minutes apart, until 3 are approved. Discussion guidelines and moderation →