Tagged articles
LLM security
36 articles · Page 1 of 1
An AI agent correctly retrieves a 'no refund' policy but still executes a $50 refund, exposing the critical gap between policy retrieval and runtime enforcement; the article argues authorization must be enforced at the tool gateway with bound decisions, not just in model context, and outlines regression tests for deny paths.
AI agentsAgent ArchitectureLLM security
0 likes · 13 min read

Despite layered defenses like input filtering, role isolation, and sandboxed execution, 2026 reports show that prompt‑injection attacks can hijack AI‑enabled browsers to click links, submit forms, leak cookies, and even trigger financial transactions, prompting a detailed analysis of attack vectors, real‑world cases, current mitigations, and emerging defenses.
AI browsersLLM securityPrompt Injection
0 likes · 12 min read

A cross‑institutional team showed that encrypted reasoning blocks in Anthropic, OpenAI and Google APIs can be replayed across sessions and models, reconstructing 315,000 blocks and leaking dozens of API keys, passwords and other sensitive artifacts, highlighting a systemic security flaw in current LLM deployments.
AI model vulnerabilityAnthropicGoogle
0 likes · 7 min read

Security researchers disclosed that Claude Opus 5 can be jailbroken with a low‑cost method that stretches the context window to 200‑300 k tokens, causing its built‑in safety guardrails to drift and allowing the model to assist in generating and improving malicious code.
Claude Opus 5LLM securitycognitive implantation
0 likes · 5 min read

WallBreaker is an open-source CLI that automates LLM red-team testing by iteratively mutating attack payloads, offering a library of research-grade techniques, a 59-to-222 transformation engine, multimodal image attacks, a HarmBench-based judge, and performance optimizations that cut token costs by 20% and boost success rates by about 30%.
AI safetyCLI toolHarmBench
0 likes · 7 min read

The article presents a three‑layer deep‑defense architecture for large language models, detailing eleven concrete methods across input filtering, execution isolation, and output verification, complete with code snippets, rule sets, and deployment guidance to prevent prompt‑injection attacks.
AI safetyAgentLLM security
0 likes · 53 min read

XG-Guard introduces a novel unsupervised graph anomaly detection framework that jointly encodes sentence- and token-level features of LLM agents, leverages theme-based anomaly scoring and covariance-based score fusion to pinpoint malicious agents in multi-agent systems, providing fine-grained explanations and enabling automatic communication isolation.
LLM securityUnsupervised Learningexplainable AI
0 likes · 9 min read

The paper reveals that prompt‑compression modules, widely used to reduce token usage in LLM agents, can be exploited to erase critical safety constraints, enabling a black‑box attack called COMA that achieves up to 71% success across multiple compressors and tasks, and proposes isolated compression as an effective defense.
Agentic AICOMALLM security
0 likes · 10 min read

The article demonstrates that vulnerability discovery depends on a flexible orchestration harness rather than on exclusive frontier LLMs, using the open‑source IronCurtain framework with commercial and open‑weight models to reproduce a 27‑year‑old OpenBSD bug and uncover new zero‑day flaws in widely used projects, while analysing costs, workflow design, and security implications.
AI orchestrationIronCurtainLLM security
0 likes · 13 min read

The article analyzes LLM jailbreak and prompt‑injection techniques—detailing five intent construction patterns, detection principles that prioritize intent over keywords, and a multi‑layered defense architecture spanning input normalization, intent analysis, generation control, and output review—to guide robust AI security.
AI safetyLLM securityPrompt Injection
0 likes · 12 min read

A new study presented at ICML 2026 reveals that the fuzzy matching used in LLM semantic caching creates an integrity vulnerability, allowing attackers to craft adversarial suffixes that cause cache‑key collisions and achieve up to 86 % response‑hijacking success on major cloud services such as AWS and Azure.
AI agentsLLM securityadversarial attacks
0 likes · 9 min read

Researchers from an international team demonstrated that the Anthropic Fable 5 model’s new safety classifier can be evaded in under five seconds with a single dialogue, exposing an internal safety collapse where agents autonomously generate harmful output during task execution, a flaw now confirmed across dozens of frontier LLMs.
AgentFable 5ISC-Bench
0 likes · 12 min read

OnlyLANs, a free AI security challenge by Just Hacking Training, lets participants jailbreak a chatbot called NetworkJohn to extract admin email, verification code, and a competitor recommendation, illustrating real‑world prompt‑injection risks highlighted in OWASP’s LLM Top‑10.
AI safetyCTFJust Hacking Training
0 likes · 3 min read

The article explains that Prompt injection is a new LLM security threat where attackers blur the line between instruction and data, outlines direct and indirect injection techniques—including command overriding, role‑play jailbreaks, encoding obfuscation, and multi‑turn attacks—and proposes a defense‑in‑depth framework with input filtering, prompt design, output validation, least‑privilege architecture, and specialized safeguards for RAG and agent scenarios.
AI safetyAgentLLM security
0 likes · 15 min read

This article experimentally evaluates a privacy‑preserving LLM security pipeline that discards raw dialogue after extracting 28 telemetry features, showing that using only 11 text‑independent signals retains about 98.5% of detection performance while reducing false‑positive rates.
LLM securityPrompt Injectionfeature engineering
0 likes · 10 min read

A systematic pressure test of Claude Code’s Auto Mode across 128 ambiguous permission scenarios reveals an 81.0% false‑negative rate and significant bypasses through Tier 2 file edits, highlighting both its partial safety benefits and critical shortcomings in autonomous code execution.
AmPermBenchAuto ModeClaude Code
0 likes · 10 min read


JavaGuideApr 14, 2026 · Artificial Intelligence
The article explains how industrial‑grade AI agents require structured prompt engineering, chain‑of‑thought reasoning, task decomposition, and a three‑layer defense (sandbox, prompt isolation, and human approval) to prevent prompt‑injection attacks, while also covering context engineering, retrieval‑augmented generation, and tool design best practices.
LLM securityagent designchain-of-thought
0 likes · 23 min read

The article debunks the hype around Claude Mythos, showing that AI‑assisted vulnerability discovery has long been a practical reality, citing VIDOC Security Lab’s findings, real‑world bug examples, the accelerating threat landscape, and recommendations for proactive, multi‑model defenses.
AI threatAI vulnerability detectionClaude Mythos
0 likes · 9 min read

This article explains the principles and common scenarios of prompt injection attacks on LLMs and provides practical defense strategies—including rule reinforcement, input filtering, output verification, and advanced techniques—to protect AI systems from malicious manipulation.
AI safetyLLM securityPrompt Injection
0 likes · 8 min read

The article presents a privacy‑first system that extracts numeric telemetry from each LLM interaction, discards raw text, and evaluates whether detection of prompt injection and jailbreak attacks remains effective, showing only a 1.4 F1‑point drop when using solely text‑independent features.
LLM securityPrompt InjectionTelemetry
0 likes · 9 min read


AntTechMar 23, 2026 · Information Security
A joint Tsinghua‑Ant Group study reveals a full‑lifecycle threat model for OpenClaw autonomous LLM agents, detailing five novel brain‑control attack vectors and proposing a five‑layer defense framework that secures the system from boot to execution.
AI safetyLLM securityautonomous agents
0 likes · 14 min read

The article analyzes how generative AI is transforming cybercrime by enabling 13 distinct attack methods—from highly personalized phishing emails and AI‑assisted malware creation to automated vulnerability hunting, deep‑fake social engineering, malicious LLMs, and attacks on AI infrastructure—highlighting recent research data and real‑world examples that illustrate the heightened speed, stealth, and accessibility of modern threats.
AI infrastructureGenerative AILLM security
0 likes · 13 min read

A recent arXiv preprint titled ‘Agents of Chaos’ details an extensive experiment where autonomous large‑language‑model agents, equipped with persistent storage, email, Discord, file system and shell access, were deployed on Fly.io VMs and subjected to red‑team attacks by twenty researchers, exposing eleven real security, privacy and governance failures.
AI riskAI safetyLLM security
0 likes · 9 min read

A comprehensive study of the OpenClaw framework reveals eleven severe security vulnerabilities in multi‑agent AI systems, ranging from over‑reactive data deletion to identity‑spoofing attacks, resource‑exhaustion loops, and covert manipulation, highlighting systemic social‑coherence failures and the need for robust agent governance.
AI agentsLLM securityOpenClaw
0 likes · 14 min read

Microsoft's AI security team introduced a lightweight scanner that detects backdoors in open‑weight large language models by leveraging three observable signals, offering a low‑false‑positive solution while highlighting the tool's methodology, limitations, and its role in extending Microsoft's AI‑focused Secure Development Lifecycle.
AI safetyLLM securityMicrosoft
0 likes · 6 min read

This article presents a systematic approach to evaluating large language model (LLM) safety by constructing an automated red‑team testing platform that measures prompt jailbreak, privacy leakage, and tool‑execution risks, defines quantitative metrics, compares commercial and open‑source models, and outlines a continuous evolution pipeline for attack samples.
AI safetyLLM securityPrompt Injection
0 likes · 20 min read

The OWASP LLM Top 10 outlines the most critical security and risk vulnerabilities in large language model applications, describing each threat—from prompt injection to model theft—its potential impact, and recommended defense principles such as secure development lifecycles, defense‑in‑depth, least‑privilege, human‑in‑the‑loop, and continuous monitoring.
AI safetyLLM securityOWASP
0 likes · 8 min read

This article presents a systematic approach to evaluating large language model security by defining threat models, categorizing attack surfaces such as jailbreak and privacy leakage, and describing an automated red‑team platform that generates, mutates, scores, and evolves adversarial prompts to continuously assess model robustness.
LLM securityPrompt InjectionRisk Assessment
0 likes · 20 min read

An extensive study reveals that twelve recent large‑language‑model defenses, including prompt‑based, adversarial‑training, filtering, and secret‑knowledge methods, are easily bypassed by a general adaptive attack framework using gradient descent, reinforcement learning, search, and human red‑team techniques, exposing critical robustness gaps.
LLM securityPrompt Injectionadaptive attacks
0 likes · 11 min read

This article reviews the latest progress in designing algorithms with strong robustness, covering adversarial examples in computer vision, novel training paradigms and certification methods, domain‑generalization techniques that achieve state‑of‑the‑art performance in medical imaging and molecular recognition, and new attack‑defense strategies for LLM jailbreak scenarios.
AI safetyLLM securityadversarial robustness
0 likes · 4 min read

A recent arXiv paper reveals that with an RTX 4070, a few hundred toxic training samples, and just one hour of fine‑tuning, attackers can embed covert advertisements into large language models like Gemini 2.5, creating cheap, undetectable AI‑driven ad platforms.
AI safetyLLM securityPrompt Injection
0 likes · 12 min read

Researchers discovered that trivial symbols such as a colon or the word “Solution” can trigger false‑positive rewards in LLM judge models, causing GPT‑4o, Claude‑4 and LLaMA‑3‑70B to fail, and proposed a robust “Master‑RM” model that eliminates these attacks.
AI robustnessLLM securityReward Modeling
0 likes · 10 min read


AntTechJun 16, 2025 · Information Security
A comprehensive study reveals that Model Context Protocol (MCP) platforms lack strict vetting, users struggle to detect malicious servers, and current large language models cannot effectively resist MCP‑level injection attacks, highlighting critical security challenges and proposing mitigation strategies.
LLM securityMCPinformation security
0 likes · 11 min read

This article explains how prompt injection attacks exploit large language models by altering their behavior through crafted inputs, outlines the major harms and attack categories—including direct, indirect, multimodal, code, and jailbreak attacks—and presents a comprehensive three‑layer defense framework covering input‑side, output‑side, and system‑level protections.
AI safetyLLM securityPrompt Injection
0 likes · 16 min read

The article examines the so‑called “grandma prompt” that tricks ChatGPT, Bing, and other LLMs into revealing Windows activation keys and even adult jokes, explains why such prompt‑injection works, and reviews past similar exploits and their mitigation attempts.
AI safetyChatGPT jailbreakLLM security
0 likes · 7 min read

The article examines the so‑called “grandma loophole”—a prompt‑injection technique that convinces ChatGPT, Bing, and other LLMs to generate Windows and Office activation keys, explores related exploits across platforms, and discusses the broader implications for AI security and ongoing mitigation efforts.
AI vulnerabilitiesChatGPTLLM security
0 likes · 7 min read
