DigestAI news desk
hot Google DeepMind launches Gemini 3.8 Live and 3.8 Live Extended Thinking models new Manchester University uses NVIDIA Earth-2 to forecast UK air pollution hot TypeSafe AI unveils Jev, a frontier model up to 400× cheaper and 200× faster new Google Cloud expands Singapore engineering centre, adds 85 AI specialists hot OpenAI uses hundreds of contractors to review ChatGPT chats new Critics label Big AI's proposed development slowdown a market lockout strategy hot Researchers Introduce 'Never Give Up' to Fix RL Reasoning Plateaus in LLMs new Jensen Huang Opposes New AI Regulations, Urges Trust in Market Forces hot Cloudflare launches Disallow AI Training to separate search indexing from model data use new Senate Drafts AI Kill Switch That Could Veto Google’s Gemini Models hot OpenAI Unveils GPT‑6 Astra: Record‑Breaking 3D Rendering, Loop‑Transformer Architecture new Anthropic projects profitability for second consecutive quarter hot OpenAI acquires AI camera software startup Glass Imaging for over $300 million new NepKANUN: RAG-based AI assistant improves access to Nepali legal information hot Mistral AI raises €3B at €21B valuation to expand European sovereign AI infrastructure new Mapping Self-Reported Personality Archetypes of 22 Large Language Models
Research updated

Typos Disrupt Prompt‑Injection Probes in LLMs, Study Finds

Researchers at arXiv have shown that ordinary typing errors—missing punctuation or simple misspellings—can dramatically weaken hidden‑state probes used to detect malicious prompts in large language models. While the models’ visible outputs remain largely unchanged, the internal readout vectors rotate by 43‑56 degrees at the typo token and decay to under 15% within about ten subsequent tokens,…

1 source primary source

Key points

  • Typos cause hidden‑state probe vectors to rotate 43‑56° and decay below 15% within ~10 tokens
  • Single‑position probe TPR drops 12 pp when three common typos are inserted
  • KV‑cache fork with downstream suffix restores 95% of lost detection performance

The team evaluated Llama‑3.1‑8B, Qwen‑3‑8B, and Gemma‑4‑E4B, finding the rotation‑and‑decay pattern consistent across them. Aggregating signals across multiple positions mitigates localized disturbances but still loses roughly 3.8 points for distributed typos. To address this, they introduced a KV‑cache fork that appends a short fixed suffix after the user message, allowing the probe to read a few tokens downstream of the error. This technique recovers 95% of the lost detection performance, outperforming perturbation‑augmented training by an order of magnitude.

The findings highlight a hidden vulnerability in current safety‑monitoring tools and suggest that even trivial user mistakes can open windows for prompt‑injection attacks, prompting a rethink of probe design for robust AI deployment.

Read the original at arXiv cs.CL · by Elad David, Max Fomin, Amit LeVi primary source Open source ↗
Topics · follow one to build your own front page
MetaAlibabaGoogle DeepMindLlama-3.1-8BQwen-3-8BGemma-4-E4B

The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us.

Comments

via GitHub Discussions

More in Research

All →

Related stories