SAINT Machines (Woosang Lim, Yeseul Chang, Jaewoo Lee) · Prompt injection defence
SPILLWAY
SPILLWAY is a runtime defence for tool-using agents that blocks prompt injection by checking where an action's target comes from, using typed tool returns rather than a language model. A deterministic sanitiser removes instruction-like text before the model reads it, and a local judge may restore refused calls only if their target is outside the removed text. Across three backbones and seven task suites on one of them, the authors report no standard templated injection redirecting an action to an attacker's target, and a mean in-scope attack success of .007 under a released reinforcement-learning attacker.
Open the source ↗Similar Research & Safety releases
- Theoretical Limits of Language Model Alignment
Apple Machine Learning Research (Lucas Monteiro Paes, Natalie Mackraz, Barry-John Theobald, Federico Danieli) · Alignment limits theory · Week 47, 2026
- Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents
Microsoft Research · Backdoor attack research · Week 47, 2026
- Anthropic files for an IPO (Form S-1)
Anthropic, PBC · IPO Filing · Week 45, 2026
- PhotoGuard
MadryLab (Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, Aleksander Madry) · Image edit protection · Week 45, 2026
- Concerning behaviour from Qwen3.8-Flash-Next
Cohere · Open-weight model safety · Week 45, 2026
- We Must Pace the Frontier
Dario Amodei · Frontier AI Pacing · Week 37, 2026
Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.