New AI Attack Hides Malicious Instructions in Normal-Looking Text to Evade Safety Filters
Check Point researchers show crafted prose hides policy-violating instructions that bypass all tested LLM gatekeepers, including GPT-4o mini and Llama Guard 3.
A new prompt-crafting technique embeds malicious payloads inside grammatical, natural-looking text without Base64, invisible Unicode, or obvious encodings, defeating lightweight pre-screening gatekeepers. In testing, all four evaluated gatekeeper models—gpt-4o-mini-2024-07-18, gpt-oss-safeguard:20b, claude-3-haiku-20240307, and llama-guard3:8b—classified the crafted wrappers as safe at a 100% bypass rate across 23 obfuscated prompts. GPT-5 Thinking in high-reasoning mode recovered and acted on the hidden instruction in 17 of 18 tests (~94.4%), often spending over a minute and multiple Python executions. Researchers recommend paraphrasing untrusted input, hardening gatekeeper policies, and applying defense-in-depth controls for agentic deployments.
- Payloads hidden in natural prose; no encodings or invisible Unicode used
- All four tested gatekeepers (GPT-4o mini, GPT-OSS Safeguard 20B, Claude 3 Haiku, Llama Guard 3) bypassed at 100%
- GPT-5 Thinking recovered embedded instructions in 17 of 18 tests (~94.4%)
- Successful recovery took over a minute of reasoning plus multiple Python scripts
- Mitigations: paraphrase untrusted text, add detection policy clauses, defense in depth
- Organizations
- Check PointOpenAIAnthropicMeta
From the original article
4 min read at gbhackers.com
A newly disclosed prompt-crafting technique can hide policy-violating instructions inside ordinary-looking English prose, allowing malicious requests to pass through lightweight LLM safety filters before being recovered and processed by a more capable downstream model. Researchers found that carefully structured prose can make the first model miss an embedded instruction entirely, while the target model invests […]
By Mayura Kathir · gbhackers.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.