LLM

2026-06-245 minSentinelOnehigh

SentinelLABS identified macOS.Gaslight, a DPRK-aligned Rust backdoor targeting macOS systems. The implant establishes a resilient C2 channel via the Telegram Bot API using AES-GCM over pinned TLS and achieves persistence via a masqueraded LaunchAgent. Notably, it embeds a 38-message prompt-injection payload designed to feed fabricated system errors to LLM-assisted triage tools, aiming to abort or corrupt automated analysis. The malware also stages a standalone Python environment to execute a credential and data stealer.

2026-06-244 minAkamaimedium

Researchers identified a critical gap in AI chatbot security where assistants leak operational context, such as tool access and boundaries, through benign reconnaissance queries. This leaked information allows attackers to bypass static model guardrails and craft highly targeted prompt injections, highlighting the need for dynamic runtime protection.

2026-06-115 minPalo Alto Networkshigh

The article introduces Behavioral Integrity Verification (BIV) to audit third-party skills for AI agents by comparing declared metadata against actual executable code and natural-language instructions. Analysis of the OpenClaw registry found that while most deviations are benign documentation errors, a critical 5% of skills contain multi-stage attack chains such as silent credential exfiltration and instruction-override hijacking.

The China-aligned threat actor UTA0388 is leveraging Large Language Models (LLMs) to conduct highly tailored, rapport-building spear-phishing campaigns targeting organizations in North America, Asia, and Europe. These campaigns deliver GOVERSHELL, a custom backdoor deployed via DLL search order hijacking, which has undergone rapid, non-iterative development across five variants to evade detection and establish persistent C2.

2026-05-135 minSophosmedium

AI agents deployed in enterprise environments are highly susceptible to indirect prompt injection attacks, enabling data theft and unauthorized actions. Security teams must adopt an 'assume breach' architecture for LLMs, focusing on blast radius reduction through agent sandboxing, credential isolation, egress restrictions, and human-in-the-loop governance.

2026-05-134 minAkamaihigh

Model Context Protocol (MCP) servers introduce a new attack surface akin to AI-native APIs, exposing organizations to protocol-level attacks, injection vulnerabilities, and authorization bypasses. Because MCP tools often use permissive validation to accommodate LLM inputs and proactively broadcast their capabilities via plain-English descriptions, attackers can easily map business logic and exploit downstream systems or trigger resource exhaustion.

2026-05-135 minRecorded Futurecritical

Recent research highlights severe security flaws in commercially available embodied AI systems, specifically Unitree humanoid and quadruped robots. Vulnerabilities including undocumented backdoors, hard-coded cryptographic keys, and unauthenticated APIs enable remote attackers to hijack devices, exfiltrate sensitive multimodal telemetry, and pivot across physical fleets via wireless interfaces.

2026-05-133 minAkamaihigh

The cybersecurity landscape is experiencing a shift towards industrialized exploitation driven by offensive AI and LLMs. These technologies act as orchestrators that rapidly discover vulnerabilities and generate exploits, necessitating defensive AI and behavioral analytics to counter machine-scale attacks.

2026-05-134 minElastic Security Labslow

Elastic Security Labs conducted research on the capabilities of Large Language Models (LLMs), specifically Claude Opus 4.6, to reverse engineer obfuscated binaries. The research demonstrates that while LLMs can defeat traditional obfuscation, novel techniques targeting LLM weaknesses—such as context window limits, budget caps, and shortcut biases—can effectively and cheaply disrupt automated static analysis pipelines.

2026-05-135 minVaronishigh

Agentic LLM browsers introduce novel architectural vulnerabilities by bridging traditional web sandboxes with highly privileged AI agents. Attackers can exploit trusted origin models via XSS or prompt injection to hijack these agents, enabling unauthorized browser control, data exfiltration, and local file access.

2026-05-134 minSentinelOneinfo

SentinelOne Labs developed a multi-agent LLM architecture using OpenClaw and Claude models to automate malware reverse engineering. By employing a serial consensus pipeline with an active rejection mandate, the system forces independent tool agents (radare2, Ghidra, Binary Ninja, IDA Pro) to cross-validate findings, significantly reducing decompiler artifacts and hallucinations.

2026-05-134 minPalo Alto Networksmedium

Unit 42 analyzed two malware samples leveraging Large Language Models (LLMs) for remote decision-making. One is a .NET infostealer using GPT-3.5-Turbo for superficial 'AI theater', while the other is a Golang dropper that uses GPT-4 to evaluate system telemetry and determine if the environment is safe to deploy a Sliver payload.

2026-05-134 minTrend Microinfo

The article outlines the emerging security risks associated with autonomous Agentic AI and presents a collaborative architectural solution between TrendAI and NVIDIA. By integrating TrendAI's governance and behavioral analysis with NVIDIA's OpenShell runtime, enterprises can safely deploy self-evolving AI agents with runtime policy enforcement and protection against AI-native threats like prompt injection.

2026-05-134 minPalo Alto Networksmedium

Unit 42 researchers demonstrated that both open and closed-source LLMs remain vulnerable to prompt jailbreaking at scale using a genetic algorithm-based fuzzing technique. By systematically generating meaning-preserving variants of disallowed requests, researchers successfully bypassed content filters and model guardrails, highlighting the fragility of current AI safety mechanisms under automated adversarial variation.

2026-05-134 minPalo Alto Networkshigh

Unit 42 researchers developed AdvJudge-Zero, an automated fuzzer that identifies stealthy prompt injection sequences to bypass AI judges. By using low-perplexity formatting tokens, attackers can manipulate LLM-based security gatekeepers into approving harmful content or corrupting training data without triggering traditional detection mechanisms.

2026-05-133 minSentinelOnelow

SentinelLabs explores the use of Large Language Models (LLMs) to automate the extraction of indicators of compromise (IOCs) and contextual data from Cyber Threat Intelligence (CTI) narratives. The research demonstrates that LLMs can accurately parse unstructured reports into structured knowledge graphs, significantly reducing processing time while highlighting the importance of custom data models, prompt optimization, and evidence-grading frameworks.

2026-05-137 minPalo Alto Networkscritical

Adversaries are actively exploiting web-based Indirect Prompt Injection (IDPI) to manipulate Large Language Models (LLMs) and AI agents. By embedding hidden or obfuscated instructions within benign web content, attackers can coerce AI systems into performing unauthorized actions such as data destruction, SEO poisoning, and bypassing content moderation when the AI processes the webpage.

2026-05-134 minSentinelOnehigh

A joint research project by SentinelLABS and Censys discovered a massive, unmanaged network of over 175,000 publicly exposed Ollama instances. Many of these self-hosted AI models possess tool-calling and vision capabilities, creating significant security risks such as resource hijacking, prompt injection, and identity laundering through residential proxy abuse.