This article explains how AgentLoop uses audit logs and local interception via LoongSuite Pilot to detect and prevent sensitive credential leaks in Qoder AI agent tasks, detailing real-world risk scenarios, investigation workflows, remediation steps, and enterprise deployment strategies.
NVIDIA's OpenShell secures AI agents by combining kernel-enforced sandbox isolation with SMT-solver-based formal verification of access policies, letting agents request permissions that are mathematically checked before human approval, while supporting Docker, Kubernetes, and MicroVM runtimes.
This article details the fundamental shift from content safety to action safety when AI agents gain real system permissions, presenting a comprehensive 10-pillar security architecture covering identity, IAM, policy engines, guardrails, sandboxes, human-in-the-loop, audit, and kill switches for production agent deployment.
NVIDIA's open-source SkillSpector scans AI agent skills for 71 vulnerability types across 17 categories using static AST/YARA/taint analysis and optional LLM semantic comparison, outputting SARIF for CI/CD gates with baseline suppression and fail-closed defaults.
This article explains how memory poisoning attacks persist across AI agent sessions, details the MINJA attack technique, and proposes auditable memory storage formats, secure write paths with trust scoring, and detection methods including integrity checks and activation monitoring to defend against persistent memory corruption.
This article dissects DeepSeek Harness's approval mechanism, showing it is a fail-closed human-machine handoff rather than a security boundary, covering its four outcome states, dual approval paths, async Cordis flow, per-package vs future-version authorization, host/client execution split, and orthogonal approval-sandbox knobs.
Anthropic’s engineering blog reveals that securing powerful AI agents like Claude requires focusing on blast radius and implementing three layered defenses—model, environment, and external content—through distinct isolation approaches, hard OS sandboxes, and practical lessons from real‑world pitfalls.
The Model Context Protocol (MCP) embeds every tool’s description into an LLM’s context window, creating a structural “Context Poisoning” vulnerability that lets malicious or bloated tool metadata hijack agent reasoning, inflate tokens, and bypass traditional input validation.
This article examines NemoClaw’s three‑layer architecture that adds kernel‑level sandboxing, policy‑driven deployment, and flexible inference routing to OpenClaw, outlines installation steps, compares it with the native OpenClaw runtime, and discusses current limitations for production use.
This article reviews recent OpenClaw security incidents, from a high‑profile email‑deletion failure caused by context compaction to supply‑chain attacks on Skills, analyzes the underlying architectural flaws of soft boundaries and missing execution‑time safeguards, and proposes a three‑layer hardening framework for AI agents.
A security analysis reveals that a high‑severity vulnerability in the open‑source AI assistant OpenClaw allows an attacker to steal the master authentication token and obtain unrestricted "god‑mode" control of the host through a single malicious link, and outlines the technical cause, attack chain, and mitigation steps.
This guide details a step‑by‑step, security‑first approach for safely experimenting with OpenClaw using a dedicated host, Tailscale private network, command whitelisting, read‑only tokens, and one‑way data flow, complete with configuration commands and emergency procedures.
AgentCore Identity, an Amazon Bedrock service, offers centralized agent identity management, secure token storage, native OAuth 2.0 support, and fine-grained access control, enabling AI agents to authenticate users, retrieve and store credentials safely, and perform actions such as scheduling Google Calendar events through a detailed end-to-end flow.
