A developer shared a Show HN demo of Semantic Overlays, small trained adapters on a frozen model that change how it perceives parts of its context. The main use case is prompt injection mitigation, claiming SOTA scores on blackbox injection benchmarks using Qwen-3.5-9B. The author invites readers to test a live demo and break the defenses.
A new method called Semantic Overlays introduces non-textual annotations to communicate span identity to language models. This mitigates prompt injection attacks, which exploit models' inability to distinguish between different input spans. The approach could enhance AI safety by addressing a key vulnerability in model serving.
A new method called Semantic Overlays introduces non-textual annotations to communicate span identity to language models. This mitigates prompt injection attacks, which exploit models' inability to distinguish between different input spans. The approach could enhance AI safety by addressing a key vulnerability in model serving.
A Medium essay argues the biggest security problem for AI agents is not what the user asks but what the agent reads. It frames prompt injection as a boundary to defend rather than a bug to fix. The thesis matters to practitioners building or securing agentic systems.
A Medium essay argues the biggest security problem for AI agents is not what the user asks but what the agent reads. It frames prompt injection as a boundary to defend rather than a bug to fix. The thesis matters to practitioners building or securing agentic systems.
ReasonGate is an open-source tool that acts as a gate to block prompt injection attacks on LLM applications. It provides explainable decisions, helping developers secure their AI systems against malicious inputs.
ReasonGate is an open-source tool that acts as a gate to block prompt injection attacks on LLM applications. It provides explainable decisions, helping developers secure their AI systems against malicious inputs.
An arXiv paper demonstrates a targeted bias injection attack against masked diffusion language models (dLLMs). Because dLLMs re-predict tokens across many denoising steps, an adversary can monitor internal activations and use a PI controller to adaptively steer a frozen model toward a chosen demographic answer. The attack substantially shifts outputs of LLaDA-8B-Instruct on bias benchmarks, and the authors call for serving-stack-level bias audits.
An arXiv paper demonstrates a targeted bias injection attack against masked diffusion language models (dLLMs). Because dLLMs re-predict tokens across many denoising steps, an adversary can monitor internal activations and use a PI controller to adaptively steer a frozen model toward a chosen demographic answer. The attack substantially shifts outputs of LLaDA-8B-Instruct on bias benchmarks, and the authors call for serving-stack-level bias audits.
A red teaming practitioner requests real-world attack examples (prompt injection, jailbreak, tool misuse) against current LLMs with strong defenses, noting that older examples like DAN for GPT-4o are outdated. The post highlights the challenge of non-deterministic LLMs and the need for up-to-date, practical datapoints.
A red teaming practitioner requests real-world attack examples (prompt injection, jailbreak, tool misuse) against current LLMs with strong defenses, noting that older examples like DAN for GPT-4o are outdated. The post highlights the challenge of non-deterministic LLMs and the need for up-to-date, practical datapoints.