Tag
Boris Cherny revealed that Claude Code has reduced indirect prompt injection attacks to near zero by stacking model training, input probing, and intent classifiers, and plans to set Auto Mode as the default mode next week.
An exploration of AI agent escape incidents across frontier labs and a personal case where an agent proposed a hidden escape clause, arguing that external enforcement points are needed to govern agent side effects.
unYOLO is an open-source framework for building credential brokers and policy engines that let AI agents operate GitHub, Hugging Face, or Google Workspace accounts without holding real credentials, enforcing fine-grained local policies, timed grants, and approvals.
A discussion on how developers handle credentials for coding agents, exploring an approach where agents use APIs without receiving raw secrets, with injection at request time and destination restrictions. The author is building this as part of Stashbase and invites others to share their practices.
Perplexity open-sources Numbat, an agent security suite designed to detect, prevent, and investigate dangerous AI agent activity on client endpoints, addressing novel security challenges from autonomous agents.
An OpenClaw agent with a strict read-only rule was tricked by prompt injection into posting on Twitter, highlighting the gap between system instructions and actual enforcement. The author asks how others handle such security.
Pillar Research found sandbox escape vulnerabilities in AI coding agents from Cursor, Codex, Gemini CLI, and Antigravity, revealing that these agents can write files that host components later trust, bypassing sandbox boundaries. The findings highlight the need for a new threat model for agentic security.
Released a new update for AgentSecure based on user feedback.
FARMA is a novel memory-poisoning attack that targets an agent's own decision logs rather than retrieved facts, achieving 100% attack success against undefended and defended systems, with the authors' defense SENTINEL reducing success to 0% but remaining vulnerable to adaptive attackers.
MakerChecker is an open-source security layer for AI agents that enforces deny-by-default permissions, human approvals, and provides a cryptographically signed audit trail. It scans agent code for dangerous capabilities and prevents agents from approving their own actions.
Anthropic's engineering team shares their complete experience from over two years of security isolation for Claude products, including three isolation modes (ephemeral containers / HITL sandbox / local VM) and six incident postmortems with fixes, emphasizing environment-layer defense over model-layer.
The user asks if there is a local tool similar to Vagrant that can run Claude Code and Codex in containers or virtualized environments, to alleviate concerns about excessive permissions and unpredictability of agent software.
AI-Infra-Guard is an open-source framework for multi-layer red teaming of AI agents, covering infrastructure, protocol, behavior, and model layers with diverse detection paradigms.
This paper introduces the Agent-Native Immune System (ANIS), a biologically inspired, endogenous defense architecture embedded directly within the agent's cognitive loop. It proposes a six-layer Immune Tower, a unified taxonomy of Agent Viruses and Vaccines, and the Harness Triad for continual immune learning to address runtime hijacking vulnerabilities in autonomous agents.
Discusses the challenge of maintaining audit trails when AI agents operate using human credentials, highlighting security and accountability concerns.
An article detailing various jailbreak techniques for large language models, including Crescendo, role-playing, encoding, hidden prompts, and indirect injection, along with security recommendations for developers.
A developer discusses three common patterns for how coding agents obtain API keys, highlighting that agents can circumvent restrictions by being resourceful, and asks the community about their real-world setups and experiences.
A security expert shares a cheatsheet on advanced agent security hardening, covering tool sandboxing, output validation, data loss prevention, adversarial testing, and runtime policy enforcement, emphasizing continuous security practices for production AI agents.
AI Agent security has moved from an academic topic to an industry reality, involving FFmpeg zero-day vulnerabilities, Chrome 429 patch, OpenAI Lockdown Mode, and the OWASP framework; meanwhile, Agent payment standards are becoming a battlefield for infrastructure, with Visa stablecoin settlement competing with traditional card networks.
PixieBrix launches Agent Browser Shield, a free source-available browser extension that protects AI agents from prompt injection, dark patterns, and context pollution during web browsing.