AICrypto
Same Bytes, Different Authority: Reserved-Token Representations in Chat-Template Prompt Injection
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 05:24:42 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 05:18:24 UTC.
Briefings
Research reveals that injecting instructions into LLM agents via reserved tokens significantly enhances attack success. Experiments show a 39 to 66 percentage point reduction in attack success when forged markers are encoded as subwords rather than reserved control tokens, with the gap persisting across multi-turn agent tasks. The effectiveness varies by model family, with Qwen3-8B showing a smaller gap due to its ability to recognize forged turns via reasoning. Instruction tuning further strengthens the model's preference for reserved markers, and standard mitigations fail to protect against attacks using these tokens.
Key points
- Injecting instructions wrapped in reserved tokens significantly enhances attack success against LLM agents, reducing attack success by 39 to 66 percentage points on open-weight families when encoded as subwords.
- Instruction tuning strengthens the model's preference for reserved markers, and the standard mitigation of encoding special tokens as subwords fails to protect against attacks using these tokens.
Context and limitations
Coverage is limited to three of four open-weight families and three of four multi-turn agent tasks in AgentDojo; results may vary across other model families or task types.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 05:24:42 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigze66klwg4fboyzinsnkh6hr2qpqv2cjlt7h25ywiiuwe7phttxe