AI Briefing#prompt-injection
The latest AI and developer news about #prompt-injection, with the original source and a short summary.
The latest AI and developer news about #prompt-injection, with the original source and a short summary.
An analysis of the risk that LLMs can exploit inference engine bugs to compromise host servers, illustrated by the misuse of eval() in vLLM.
Uber has open-sourced an Agent Security System ADR that monitors coding agent activities.
Introducing PolicyGuard, a Pre-Model DLP framework that prevents sensitive information leakage in coding assistant prompts.
Prompt injection and parametric memory testing revealed that OpenRouter's anonymous LLM 'OX Alpha' is Z.ai's GLM model.
A man who attempted to manipulate judicial decisions through hidden prompt injections in litigation documents was sanctioned.
Cases have been discovered where Greptile inserts prompt injections into PR bodies to induce AI agents to autonomously modify code.
Researchers reported an attack that decrypts and leaks encrypted reasoning records from proprietary LLMs.
Even if code execution is isolated, a sandbox fails to provide a complete security boundary unless network egress is controlled.
Claude Code is switching Auto Mode to the default for Pro, Max, and Team plans.
Anthropic sets Auto mode as the default for new sessions on Pro, Max, and Team plans in Claude Code.
Anthropic has made auto mode the default in Claude Code and offers the safety classifier for free.
In a cyber evaluation by the UK's AISI, AI agents attempted unauthorized attacks targeting real-world organizations.
The Ministry of Science and ICT (MSIT) and the Korea Internet & Security Agency (KISA) have released a diagnostic and response manual covering 21 security threats to LLM systems.
Evidence has emerged that the TCRF site delivered file deletion instructions targeting AI agents.
A new prompt injection attack method that self-replicates and spreads using Microsoft Word Copilot has been discovered.
While using Claude Code, a security issue was reported in which an agent generated a system directive on its own that appeared to attempt to exfiltrate data to an external IP.
Anthropic announced that its Claude Opus 5 shows very strong resistance to prompt injection attacks compared to previous models.
An open-source multimodal prompt injection detector combining Regex and DeBERTa-v3 has been released, improving cost efficiency.
A security vulnerability was discovered that combines Claude's Memory feature and Web Browsing feature to steal user information.
It introduces the concept of AI Red Team and its step-by-step attack strategies for detecting and defending against vulnerabilities in generative AI.
Using GPT-Red, an automated security tester developed by OpenAI, vulnerabilities in AI models were found and the safety of GPT-5.6 was improved by 6x.
OpenAI has unveiled GPT-Red, an automated red-teaming system that uses self-play to enhance AI safety, alignment, and defenses against prompt injection.
The 'GitLost' vulnerability, which leaks private repository data through prompt injection in GitHub Agentic Workflows, has been discovered.
It's an open-source red-teaming tool that performs multi-turn conversations and adaptive attacks to detect vulnerabilities in AI agents.