CVE-2026-24222

NVIDIA NeMoClaw contains a vulnerability in the sandbox environment initialization component, where a remote attacker could cause improper access control by sending prompt-injected content that causes the agent to read and exfiltrate host environment variables not properly restricted during sandb...

Logic-to-Code Execution via Indirect Prompt Injection

This document explores a critical architectural vulnerability in Large Language Model LLM implementations, specifically within Command Line Interface CLI tools and automated agentic workflows. The research demonstrates how the absence of separation between the control plane instructions and the...

CVE-2026-24222: Exposure of Sensitive System Information to an Unauthorized Control Sphere

NVIDIA NeMoClaw contains a vulnerability in the sandbox environment initialization component, where a remote attacker could cause improper access control by sending prompt-injected content that causes the agent to read and exfiltrate host environment variables not properly restricted during sandb...

An Empirical Security Evaluation of LLM-Generated Cryptographic Rust Code

Developers and organizations are using Large Language Models LLMs to generate security-critical code more frequently than ever, including cryptographic solutions for their products. This study presents an empirical evaluation of cryptographic security in 240 Rust code samples for two crypto...

PT-2026-35756

NVIDIA NeMoClaw contains a vulnerability in the sandbox environment initialization component, where a remote attacker could cause improper access control by sending prompt-injected content that causes the agent to read and exfiltrate host environment variables not properly restricted during sandb...

AgentVisor: Defending LLM Agents against Prompt Injection Via Semantic Virtualization

Large Language Model LLM agents are increasingly used to automate complex workflows, but integrating untrusted external data with privileged execution exposes them to severe security risks, particularly direct and indirect prompt injection. Existing defenses face significant challenges in balanci...

Evaluation of Prompt Injection Defenses in Large Language Models

LLM-powered applications routinely embed secrets in system prompts, yet models can be tricked into revealing them. We built an adaptive attacker that evolves its strategies over hundreds of rounds and tested it against nine defense configurations across more than 20,000 attacks. Every defense tha...

Insufficient Granularity of Access Control

Overview openclaw is a 🦞 OpenClaw — Personal AI Assistant Affected versions of this package are vulnerable to Insufficient Granularity of Access Control via insufficient access control in the gateway config.patch and config.apply processes. An attacker can modify protected operator settings by...

Evaluating Jailbreaking Vulnerabilities in LLMs Deployed As Assistants for Smart Grid Operations: A Benchmark against NERC Standards

The deployment of Large Language Models LLMs as assistants in electric grid operations promises to streamline compliance and decision-making but exposes new vulnerabilities to prompt-based adversarial attacks. This paper evaluates the risk of jailbreaking LLMs, i.e., circumventing safety alignmen...

Semantic Denial of Service in LLM-Controlled Robots

Safety-oriented instruction-following is supposed to keep LLM-controlled robots safe. We show it also creates an availability attack surface. By injecting short safety-plausible phrases 1-5 tokens into a robots audio channel, an adversary can trigger the models safety reasoning to halt or disrupt...

PT-2026-37322

Affected Packages / Versions - Package: openclaw npm - Affected versions: 2026.4.20 - Patched version: 2026.4.20 Impact The agent-facing gateway config.patch / config.apply guard did not cover several operator-trusted settings, including sandbox policy, plugin enablement, gateway auth/TLS, hook...

PT-2026-37320

Affected Packages / Versions - Package: openclaw npm - Affected versions: 2026.4.20 - Patched version: 2026.4.20 Impact Output from webhook-triggered isolated cron agent runs could be queued into the main session awareness stream without trusted: false. That made the event render as a trusted...

k8sGPT has Prompt Injection through its k8sGPT-Operator

Summary In the auto-remediation pipeline, objecttoexecution.go was deserializing the AI-generated YAML directly into a Deployment object, but there was lack of validation from the original Deployment object. Details This issue was fixed after coordination with Alex Jones. PoC To minimize the...

GHSA-RP7V-4384-HFRP k8sGPT has Prompt Injection through its k8sGPT-Operator

Summary In the auto-remediation pipeline, objecttoexecution.go was deserializing the AI-generated YAML directly into a Deployment object, but there was lack of validation from the original Deployment object. Details This issue was fixed after coordination with Alex Jones. PoC To minimize the...

CVE-2026-41318

AnythingLLM prior to v1.12.1 is vulnerable to stored DOM-based XSS via an unsafe image rendering rule and unsanitized chart captions in the Chartable component. The vulnerability arises because renderMarkdown(...) output is sanitized in all call sites except Chartable, where LLM-generated caption...

PT-2026-37208

Name of the Vulnerable Software and Affected Versions @google/gemini-cli versions prior to 0.39.1 run-gemini-cli affected versions not specified Description Issues exist in the handling of workspace trust and tool allowlisting when running in headless mode, such as within CI/CD environments. In...

PT-2026-37176

Name of the Vulnerable Software and Affected Versions LiteLLM versions 1.80.5 through 1.83.6 Description The 'POST /prompts/test' endpoint accepts user-supplied prompt templates and renders them without sandboxing. An authenticated user with a valid proxy API key can provide a crafted template to...

Anthropic Claude Code < 2.1.64 Sandbox Escape via Symlink Following (CVE-2026-39861)

The version of Anthropic Claude Code installed on the remote host is prior to 2.1.64. It is, therefore, affected by a sandbox escape vulnerability. - Claude Code's sandbox did not prevent sandboxed processes from creating symlinks pointing to locations outside the workspace. When Claude Code...

PT-2026-36134

Impact The POST /prompts/test endpoint accepted user-supplied prompt templates and rendered them without sandboxing. A crafted template could run arbitrary code inside the LiteLLM Proxy process. The endpoint only checks that the caller presents a valid proxy API key, so any authenticated user cou...

📄 OpenClaw 2026.3.13 MEDIA Protocol File Disclosure

This Python script is a security exploitation tool targeting the OpenClaw system integrated with Discord. It attempts to exfiltrate sensitive files from a victim environment by abusing a MEDIA: prompt injection mechanism...