reverse-captcha-eval

Reverse CAPTCHA: Evaluación de la susceptibilidad de los LLM a la inyección invisible de instrucciones Unicode Un marco de evaluación que prueba si los modelos de lenguaje grandes siguen instrucciones codificadas en Unicode invisible incrustadas en texto que de otro modo parece normal. Mientras q...

RAISED: Self-Distillation for Robustness to Prompt Injection in LLM Agents

Tool-using language-model agents are vulnerable to indirect prompt injection because they must act on untrusted external content. Existing training-time defenses can reduce attack success rates, but often at the cost of general capabilities. We show that training-based defenses induce substantial...

Configuration, Not Conscience: A Large-Scale Empirical Study of LLM System Prompts

Leaked system prompts are often treated as windows into the hidden values of commercial language models, yet their composition is rarely studied at scale. We analyze a merged corpus of 407 leaked, reconstructed, or officially published system prompts from 62 vendors across four community...

AURA-Eval: Evaluation Framework for Acting under Risk Awareness in LLM Agent Trajectories

LLM agents operate in workflows where unsafe actions can have real consequences. Existing safety evaluations often reduce behavior to a single score, obscuring risk recognition, pre-action detection, and safe task completion when a safe solution exists. We introduce AURA-Eval, a framework combini...

Graph Is the Verifier: Agentic Reinforcement Learning for Interprocedural Vulnerability Detection

Real-world vulnerabilities often span multiple functions, yet most learning-based detectors classify each function in isolation: on a sample of real CVEs, we find that 71.7% of vulnerable functions require evidence from outside the function to be classified correctly. Agentic reinforcement learni...

CVE-2026-58500

MCP Appium is an MCP server that provides AI assistants with tools to automate mobile app testing on Android and iOS. In versions prior to 1.85.10, the createLocatorGeneratorUI function interpolates attacker-controlled element attributes — text, content-desc, resource-id, and locator selector...

CVE-2026-61439 PraisonAI before 4.6.78 Prompt Injection Defense Bypass

PraisonAI versions before 4.6.78 contain a prompt injection defense misconfiguration where the block threshold defaults to CRITICAL severity, allowing HIGH-level threats to pass through unblocked. Attackers can submit single-vector prompt injection attacks such as instruction overrides or financi...

The Cost of Non-Compliance: Why AI Governance Is the New Enterprise Imperative

AI governance helps enterprises control tool use, reduce compliance risk, protect customer data, and avoid fines as teams adopt AI faster than policy...

Incorrect Authorization

Overview praisonai is a PraisonAI TypeScript AI Agents Framework - Node.js, npm, and Javascript AI Agents Framework Affected versions of this package are vulnerable to Incorrect Authorization due to the approval callback onToolCall being invoked only after the execution of tools in the AgentLoop...

PT-2026-50534

Name of the Vulnerable Software and Affected Versions Network-AI versions prior to 5.7.2 Description The MCP SSE server allows unauthenticated cross-origin MCP tool invocation because the server defaults to an empty secret and the isAuthorized function returns true when the secret is empty. While...

From AI-Generated Content to Agentic Action: Security and Safety Threats in Generative AI

Generative AI systems are increasingly used not only to produce content but also to retrieve data, invoke tools, and execute actions. This work examines the security and safety implications of that shift across content-level, model-level, and agentic threats. We analyze how attacker access...

Oracle Poisoning: Corrupting Knowledge Graphs to Weaponise AI Agent Reasoning

We define Oracle Poisoning, an attack class in which an adversary corrupts a structured knowledge graph that AI agents query at runtime via tool-use protocols, causing incorrect conclusions through correct reasoning. Unlike prompt injection, Oracle Poisoning manipulates the data agents reason ove...

2026: The Year of AI-Assisted Attacks

On December 4, 2025, a 17-year-old was arrested in Osaka under Japan’s Unauthorized Access Prohibition Act. The young man had run malicious code to extract the personal data of over 7 million users of Kaikatsu Club, Japan's largest internet cafe chain. When asked, the young man shared his...

Your LLM Agent Can Leak Your Data: Data Exfiltration Via Backdoored Tool Use

Tool-use large language model LLM agents are increasingly deployed to support sensitive workflows, relying on tool calls for retrieval, external API access, and session memory management. While prior research has examined various threats, the risk of systematic data exfiltration by backdoored...

AgentWatcher: A Rule-Based Prompt Injection Monitor

Large language models LLMs and their applications, such as agents, are highly vulnerable to prompt injection attacks. State-of-the-art prompt injection detection methods have the following limitations: 1 their effectiveness degrades significantly as context length increases, and 2 they lack...

Addressing the OWASP Top 10 Risks in Agentic AI with Microsoft Copilot Studio

Agentic AI is moving fast from pilots to production. That shift changes the security conversation. These systems do not just generate content. They can retrieve sensitive data, invoke tools, and take action using real identities and permissions. When something goes wrong, the failure is not limit...

CVE-2026-30856

WeKnora is an LLM-powered framework designed for deep document understanding and semantic retrieval. Prior to version 0.3.0, a vulnerability involving tool name collision and indirect prompt injection allows a malicious remote MCP server to hijack tool execution. By exploiting an ambiguous naming...

Reverse CAPTCHA: Evaluating LLM Susceptibility to Invisible Unicode Instruction Injection

We introduce Reverse CAPTCHA, an evaluation framework that tests whether large language models follow invisible Unicode-encoded instructions embedded in otherwise normal-looking text. Unlike traditional CAPTCHAs that distinguish humans from machines, our benchmark exploits a capability gap: model...

AJAR: Adaptive Jailbreak Architecture for Red-Teaming

As Large Language Models LLMs evolve from static chatbots into autonomous agents capable of tool execution, the landscape of AI safety is shifting from content moderation to action security. However, existing red-teaming frameworks remain bifurcated: they either focus on rigid, script-based text...

The Evolution of Agentic AI in Cybersecurity: From Single LLM Reasoners to Multi-Agent Systems and Autonomous Pipelines

Cybersecurity has become one of the earliest adopters of agentic AI, as security operations centers increasingly rely on multi-step reasoning, tool-driven analysis, and rapid decision-making under pressure. While individual large language models can summarize alerts or interpret unstructured...