LLM Security

Anthropic Mythos LLM Scans Federal Software for Vulnerabilities

ATLAS OWASP MEDIUM ▲ 6.5 SecurityWeek Jul 08, 2026

CISA's Attack Surface Evaluation team is reportedly leveraging Anthropic's 'Mythos' model to scan federal government software for security vulnerabilities, representing a significant expansion of AI-assisted offensive security tooling in critical infrastructure defence. The deployment raises important questions about the trustworthiness of LLM-driven vulnerability assessment, potential for model-induced false negatives, and the security of the AI pipeline itself when applied to sensitive government codebases. This marks one of the most prominent known uses of a commercial LLM in an active U.S. government cyber defence role.

Tencent Releases Hy3 295B Open-Source Model with 256K Context

FIRST LOOK ATLAS OWASP MEDIUM ▲ 5.5 Simon Willison Jul 07, 2026

Tencent has released Hy3, a 295B-parameter Mixture-of-Experts open-source model under Apache 2.0, featuring 256K context length and temporarily available for free inference via OpenRouter. The model's large context window, open weights, and Chinese provenance expand the attack surface for defenders managing LLM supply chains, jailbreak campaigns, and influence operations. Security teams should treat this as another high-capability open-weight model requiring the same scrutiny applied to comparable releases from Mistral or Meta.

AWS Launches Multi-Turn RL for Amazon Nova

FIRST LOOK ATLAS OWASP HIGH ▲ 7.2 AWS Machine Learning Blog Jul 07, 2026

AWS has released a production-grade, event-driven multi-turn reinforcement learning training infrastructure for Amazon Nova models on SageMaker HyperPod, enabling enterprises to train agents that learn tool orchestration, error recovery, and sequential decision-making at scale. This materially expands the attack surface by introducing complex reward-routing pipelines, ephemeral compute provisioning, and environment-facing reward workers as new targets for poisoning and manipulation. Defenders must scrutinise the trust boundaries between the Nova Forge SDK, ECS reward workers, and HyperPod training pods, as a compromised reward signal can silently shape model behaviour across entire interaction sequences.

OfficeCLI Brings Microsoft Office Automation to AI Agents

FIRST LOOK ATLAS OWASP HIGH ▲ 7.8 HN AI Security Jul 07, 2026

OfficeCLI is an open-source, single-binary tool that enables AI agents to programmatically read, write, and automate Microsoft Word, Excel, and PowerPoint files without requiring a local Office installation. This dramatically expands the file-system attack surface for agentic AI systems, enabling prompt injection via document content, automated exfiltration of sensitive Office files, and weaponisation of documents as a persistent injection vector. Defenders operating AI agent pipelines that touch file systems must now treat any Office document as a potential adversarial input channel.

Prompt Injection Attacks Manipulate AI Crypto Agents

ATLAS OWASP HIGH ▲ 8.5 SecurityWeek Jul 07, 2026

Researchers identified two active campaigns embedding indirect prompt injection payloads in malicious websites to manipulate autonomous AI agents into executing unauthorised cryptocurrency transactions. The attacks exploit the growing deployment of agentic AI systems that browse the web and take real-world actions with minimal human oversight. This represents a concrete, financially motivated escalation of prompt injection from data exfiltration to direct fund theft.

SkillCloak Bypasses AI Agent Skill Scanners at 90% Rate

ATLAS OWASP HIGH ▲ 8.5 The Hacker News Jul 07, 2026

Researchers at Hong Kong University of Science and Technology have demonstrated that static scanners used to vet malicious AI agent 'skills' — modular add-ons for agents like Claude Code and OpenAI Codex — can be systematically bypassed using a tool called SKILLCLOAK. The technique leverages either character-substitution obfuscation or self-extracting packing into scanner-ignored directories like .git/, achieving evasion rates above 90% across all eight tested scanners. The same research team also developed SKILLDETONATE, a runtime behavioral sandbox that catches most of the threats static analysis misses.

Amazon Q Extension Credential Theft via MCP Injection

ATLAS OWASP HIGH ▲ 8.5 Dark Reading Jul 05, 2026

A vulnerability in the Amazon Q Visual Studio Code extension allows adversaries to plant malicious repositories that execute arbitrary code and exfiltrate cloud credentials. The flaw highlights escalating risks associated with Model Context Protocol (MCP) integrations embedded within AI-powered developer tools. This attack vector represents a growing threat surface as AI coding assistants gain privileged access to developer environments and cloud infrastructure.

Alibaba and Baidu Launch LLMs With US-Level Capabilities

FIRST LOOK ATLAS OWASP HIGH ▲ 6.2 Dark Reading Jul 04, 2026

Two newly released large language models from Chinese AI firms have reached capability parity with leading US frontier models, expanding the global pool of powerful AI available to both commercial and adversarial users. For defenders, this development broadens the asymmetry between attackers — who gain access to capable, potentially less-restricted models — and defenders, who must now account for threats generated by a wider set of model providers. Security teams should anticipate increased use of these models for offensive tasks such as phishing content generation, vulnerability research automation, and social engineering at scale.

Langflow LLM Agents Exploited for Ransomware Delivery

ATLAS OWASP CRITICAL ▲ 9.2 SecurityWeek Jul 04, 2026

A documented ransomware attack leveraged agentic AI infrastructure — specifically the Langflow LLM orchestration platform — to automate multi-stage intrusion chains combining known exploitation techniques with real-time LLM reasoning. This marks a significant escalation in threat actor capability, demonstrating that agentic AI can serve as an autonomous attack coordinator rather than merely an assistant. Security teams running self-hosted AI orchestration platforms now face an expanded attack surface where the AI layer itself can be both the entry point and the execution engine.

Microsoft Copilot MCP Tool Poisoning Enables Data Exfiltration

ATLAS OWASP HIGH ▲ 9.1 The Hacker News Jul 04, 2026

Microsoft researchers have demonstrated how attackers can embed hidden instructions inside MCP tool descriptions to covertly redirect AI agents into exfiltrating sensitive business data. Because each individual action the agent takes appears legitimate — using approved tools and the user's own permissions — default security controls generate no alerts. The attack exploits a fundamental design tension in MCP: tool descriptions simultaneously carry operational instructions and attacker-controlled data, collapsing a critical trust boundary.

Agentjacking: Prompt Injection via Malicious Bug Reports

ATLAS OWASP HIGH ▲ 8.2 Dark Reading Jul 04, 2026

A technique dubbed 'agentjacking' exploits the inability of AI coding agents to distinguish between legitimate content and embedded instructions, allowing attackers to hijack agent behaviour through maliciously crafted bug reports. The attack represents a scalable, low-barrier prompt injection vector targeting developer workflows that rely on autonomous AI agents. As AI coding assistants gain broader adoption and elevated system permissions, this class of attack poses a significant risk to software supply chain integrity.

Cursor IDE DuneSlide Zero-Click Prompt Injection RCE

ATLAS OWASP CRITICAL ▲ 9.2 SecurityWeek Jul 04, 2026

A set of vulnerabilities dubbed 'DuneSlide' in the Cursor AI code editor allow attackers to conduct zero-click prompt injection attacks that escape the application's sandbox and execute arbitrary code at the operating system level. The flaws represent a critical escalation of AI-native attack surface risks, targeting developers who rely on AI-assisted coding environments. Because exploitation requires no user interaction, the attack chain is particularly dangerous in supply chain and watering-hole scenarios.

Browser Ransomware via File System Access API: DeepSeek

ATLAS OWASP HIGH ▲ 8.5 Check Point Research Jul 03, 2026

Check Point Research demonstrates how DeepSeek's lower refusal rates allowed researchers to transform an LLM-hallucinated malware concept into a practical browser-native ransomware technique targeting Android photo directories via the File System Access API. The attack requires no native payload, APK installation, or root access — only social engineering to obtain a legitimate browser permission prompt. This research highlights how frontier AI models with weaker safety controls can independently design novel attack paths not yet seen in real-world campaigns.

CVE-2026-50548: Cursor IDE Prompt Injection RCE

ATLAS OWASP CRITICAL ▲ 9.2 The Hacker News Jul 03, 2026

Two critical vulnerabilities (CVE-2026-50548 and CVE-2026-50549) in the Cursor AI code editor allow prompt injection attacks delivered via MCP services or web search results to escape the editor's terminal sandbox and execute arbitrary commands on a developer's machine without any user interaction. Both flaws abuse the sandbox's write-permission logic — one through a misconfigured working directory parameter, the other through a symlink-resolution fallback — ultimately allowing overwrite of the sandbox helper binary itself. The attack surface is significant given Cursor's reported adoption across more than half of Fortune 500 companies; all versions prior to 3.0 remain vulnerable.

Anthropic Releases Claude-Real-Video for Local Video Analysis

FIRST LOOK ATLAS OWASP MEDIUM ▲ 6.8 HN AI Security Jul 03, 2026

claude-real-video is an open-source, MIT-licensed Python library that extracts scene-change frames, deduplicates images, and transcribes audio from any video URL or local file, then packages the result as a folder any LLM can consume — all processed locally without cloud upload. For defenders, this dramatically expands the multimodal prompt injection surface by enabling adversaries to embed malicious instructions inside video content that LLM pipelines will now ingest and act upon. Security teams building or deploying LLM agents with video-processing capabilities must treat video content as an untrusted, potentially adversarial input channel.

LLM Security

Stay ahead of the threat.