Security

In this tutorial, we explore how to design production-grade safety for LLM-based applications using the NeMo Guardrails framework. We move beyond simple prompt filtering to implement a layered architecture, featuring deterministic PII redaction, retrieval filtering, output masking, and policy-based tool gating. By integrating stateful multi-turn evaluation and detailed activation tracing, we demonstrate how to build an auditable, secure, and cost-effective AI assistant capable of managing sensitive financial interactions while maintaining strict compliance standards

Anthropic has moved its most cyber-capable model into a product security teams can switch on themselves. Claude Security scans now run on Claude Mythos 5, in public beta for Claude Enterprise customers with no separate model add-on. The scan connects to a GitHub repository, traces data flows across files, and returns findings with a CWE category, confidence and severity ratings, and a suggested patch. The design point is packaging: users receive a scan result rather than a prompt box, so the model that finds vulnerabilities cannot be steered into writing exploits.

Mistral AI has released Shieldstral 1.0 3B, an open-weights, policy-adaptive multimodal safety classifier that frames content moderation as a single yes/no question instead of a fixed harm taxonomy. Operators supply the policy as a plain-language query at inference time and get back a calibrated safety score from one forward pass — no retraining required to re-target the model. Built on Ministral-3-3B-Base-2512 with a Pixtral vision encoder and trained on roughly 54.1M samples, it reports 84.9% average F1 on text safety (matching GPT-OSS-Safeguard-20B), 83.8% on multimodal safety, and 91.3% on Mistral's adaptability benchmark — while fitting in 16GB of VRAM under an Apache 2.0 license.

AI agents, MCP servers, and LLM apps break the core AppSec assumption that applications do what their code says. This guide walks through a practical see-fix-protect framework: a five-layer agentic AI attack surface map, a 12-point misconfiguration checklist, an evidence-based triage matrix, runtime guardrails, and system prompt hardening — plus a maturity self-assessment aligned to NIST AI RMF, OWASP AIMA, ISO/IEC 42001, and the EU AI Act.

Cogent AI team released Cogent VR-1, a reasoning model post-trained specifically for cybersecurity rather than picking up cyber capability as a side effect of...

Microsoft AI has released MAI-Cyber-1-Flash, its first model built specifically for cyber defense. It is a 137B total, 5B active sparse MoE fine-tune of MAI-Code-1-Flash with a 256k context window. The model does not ship as a standalone endpoint — it runs inside MDASH, Microsoft's multi-model agentic scanning harness, where it handles up to 90% of tasks and pushes the system to 95.95% on CyberGym.

Sakana AI has released Fugu-Cyber, a security-tuned endpoint on its Fugu orchestration model. It reports 86.9% on CyberGym and 72.1% on CTI-REALM, edging past GPT-5.5-Cyber and Claude Mythos Preview. Access is gated behind manual approval, a defensive-use policy, and the Token Plan. Here is what the numbers actually mean.

OpenAI disclosed that its own models breached Hugging Face's production infrastructure while taking a public security benchmark. The models were not attacking a target — they were optimizing a score. Here is the mechanism, what the ExploitGym data showed two months earlier, and which widely repeated claims about the incident are not actually confirmed.

Anthropic has released the Claude Security plugin for Claude Code in beta. The plugin runs a multi-agent vulnerability scan of a repository from inside...

Cisco Foundation AI has released Antares, a family of small language models trained to pinpoint where known vulnerabilities live inside a codebase. Antares-1B reaches 0.209 File F1 on the new Vulnerability Localization Benchmark, above GLM-5.2 at 753B parameters and Gemini 3 Pro. The untrained Granite 4.0 checkpoints score near zero under the same protocol, so post-training supplies almost all of the capability. A full 500-task sweep runs in roughly 13 minutes on a single H100 for under a dollar, against $141 for GPT-5.5.

OpenAI trained GPT-Red, an internal-only attacker model, using self-play reinforcement learning against a population of defender LLMs. It beat human red-teamers 84% to 13% on a replicated indirect prompt injection arena, found a novel "Fake Chain-of-Thought" attack class, and cut GPT-5.6 Sol's failures 6x on OpenAI's hardest direct injection benchmark. OpenAI concedes it still struggles with multi-turn and image-based attacks.

We build a Colab-ready PyGraphistry workflow for interactive graph analytics on enterprise access data. We generate a synthetic dataset of users, devices, IPs, services, roles, and geos, then convert it into nodes and edges. We enrich the graph with risk scores, centrality metrics, community detection, Isolation Forest anomaly scores, and UMAP layout embeddings. We then bind the graph in PyGraphistry and produce local PyVis visualizations for full, ego, and high-risk views.

shutdown followed a US government export control directive citing national security authorities. All other Anthropic models, including Opus 4.8, remain available.

In this tutorial, we explore the ClawHub Security Signals dataset to see how scanners assess AI skills. We load the data from the Hugging Face Parquet conversion and inspect verdicts, scanner outputs, and severity labels. We measure how VirusTotal, static analysis, and SkillSpector overlap and disagree using Jaccard scores and Cohen's kappa. Finally, we combine SKILL.md text with scanner signals to train a logistic regression model for ClawScan verdicts.

This tutorial walks through NVIDIA garak as an end-to-end framework for defensive LLM red-teaming. It covers setup, plugin discovery, dry runs, real-model scans on a Hugging Face generator, and multi-probe evaluations. The workflow then analyzes safety scores and attack success rates, inspects flagged outputs, and extends garak with a custom probe and detector. It closes by exporting results in AVID format for structured vulnerability

Perplexity has open-sourced Bumblebee, an internal security tool it uses to protect the developer systems behind its search product, Comet, and Computer. Bumblebee is a read-only inventory collector for macOS and Linux developer endpoints. It scans npm, PyPI, Go modules, MCP configs, editor extensions, and browser extensions — without invoking any package manager or running any code.

In this tutorial, we build a realistic Zero-Trust network simulation by modeling a micro-segmented environment as a directed graph and forcing every request to...

Fastino Labs has released GLiGuard, a 300M parameter open-source safety moderation model that evaluates four safety tasks — prompt safety, jailbreak strategy detection, harm category classification, and refusal detection — in a single forward pass. Built on an encoder architecture rather than the decoder-only design used by most guardrail models, GLiGuard achieves up to 16x higher throughput and 16.6x lower latency than current state-of-the-art models, while matching or exceeding the accuracy of models 23 to 90 times its size across nine safety benchmarks. Model weights are available under the Apache 2.0 license on Hugging Face.

OpenAI on just launched Daybreak, a cybersecurity initiative that combines the company's frontier AI models with Codex Security, its coding-focused agentic system, and a...

In this tutorial, we explore how FLARE-FLOSS helps us recover hidden and obfuscated strings from a Windows PE file. We begin by setting up...

OpenAI's Privacy Filter Is a 1.5B-Parameter PII Detector Built on a Distilled Decoder — And It Runs in Your Browser

OpenAI is moving from a limited pilot to broad deployment of its Trusted Access for Cyber program, introducing GPT-5.4-Cyber — a variant of GPT-5.4 fine-tuned to be cyber-permissive — for thousands of vetted defenders.

In this tutorial, we build a workflow that combines Magika’s deep-learning-based file type detection with OpenAI’s language intelligence to create a practical and insightful...

As Generative AI matures, so do the threats against it. AI Red Teaming has evolved from a niche security practice into a regulatory requirement. Our 2026 guide breaks down the top 19 tools—including Mindgard, Garak, and Microsoft’s PyRIT—to help security teams identify vulnerabilities like data leakage and bias before they reach production.

In this tutorial, we build a robust, multi-layered safety filter designed to defend large language models against adaptive and paraphrased attacks. We combine semantic...

In this tutorial, we build an advanced, multi-turn crescendo-style red-teaming harness using Garak to evaluate how large language models behave under gradual conversational pressure....

In this tutorial, we demonstrate a realistic data poisoning attack by manipulating labels in the CIFAR-10 dataset and observing its impact on model behavior....

OpenAI has released a research preview of gpt-oss-safeguard, two open weight safety reasoning models that let developers apply custom safety policies at inference time....

Table of contentsOverviewWhat MCP standardizes?Normative authorization controlsWhere MCP supports security engineering in practice ?Case study: the first malicious MCP serverUsing MCP to structure red-team...

When deploying AI into the real world, safety isn’t optional—it’s essential. OpenAI places strong emphasis on ensuring that applications built on its models are...

Can your AI security stack profile, reason, and neutralize a live security threat in ~220 ms—without a central round-trip? A team of researchers from...

Can safety keep up with real-time LLMs? Alibaba’s Qwen team thinks so, and it just shipped Qwen3Guard—a multilingual guardrail model family built to moderate...

In this tutorial, we introduce a Jailbreak Defense that we built step-by-step to detect and safely handle policy-evasion prompts. We generate realistic attack and...

In this tutorial, we’ll explore how to implement OAuth 2.1 for MCP servers step by step. To keep things practical, we’ll build a simple...

Machine learning (ML) is transforming industries, powering innovation in domains as varied as financial services, healthcare, autonomous systems, and e-commerce. However, as organizations operationalize...

The AI security arms race is in full swing. As cyber threats grow more sophisticated, organizations are reimagining defense strategies—with artificial intelligence taking center...

In this tutorial, we’ll explore how to test an OpenAI model against single-turn adversarial attacks using deepteam. deepteam provides 10+ attack methods—like prompt injection, jailbreaking,...

Estimated reading time: 6 minutes Table of contentsIntroductionWhat Is a Proxy Server?Technical Architecture:Key Functions (2025):Types of Proxy ServersDescriptions:Key Use Cases in 2025Emerging Trends in Proxy...

As we move through 2025, artificial intelligence (AI) is fundamentally reshaping secure browsing and Virtual Private Network (VPN) technologies. The explosion of sophisticated cyber...

As large language models (LLMs) evolve from simple text generators to agentic systems —able to plan, reason, and autonomously act—there is a significant increase...

Cybersecurity has become a significant area of interest in artificial intelligence, driven by the increasing reliance on large software systems and the expanding capabilities...

The Model Context Protocol (MCP) represents a powerful paradigm shift in how large language models interact with tools, services, and external data sources. Designed...

As AI agents become more autonomous—capable of writing production code, managing workflows, and interacting with untrusted data sources—their exposure to security risks grows significantly....

Large language models (LLMs) have become an integral part of various applications, but they remain vulnerable to exploitation. A key concern is the emergence...

Bagel is a novel AI model architecture that transforms open-source AI development by enabling permissionless contributions and ensuring revenue attribution for contributors. Its design...

Dense embedding-based text retrieval has become the cornerstone for ranking text passages in response to queries. The systems use deep learning models for embedding...

Code generation AI models (Code GenAI) are becoming pivotal in developing automated software demonstrating capabilities in writing, debugging, and reasoning about code. However, their...

Large Language Models (LLMs) have gained significant attention in recent years, but they face a critical security challenge known as prompt leakage. This vulnerability...

Large Language Models (LLMs) are vulnerable to jailbreak attacks, which can generate offensive, immoral, or otherwise improper information. By taking advantage of LLM flaws,...

Artificial Intelligence (AI) safety has become an increasingly crucial area of research, particularly as large language models (LLMs) are employed in various applications. These...

Large Language Models (LLMs) have become increasingly important in cybersecurity, particularly in their application to secure coding practices. As these AI-driven models can generate...

Cybersecurity is a fast-paced area wherein knowledge and mitigation of threats are most necessary. In this respect, the attack graph is one tool that...

LLMs have shown impressive abilities, generating contextually accurate responses across different fields. However, as their capabilities expand, so do the security risks they pose....

Ensuring data privacy and security during computational processes presents a significant challenge, particularly when using cloud services. Traditional encryption methods require data to be...

Hackers finding a way to mislead their AI into disclosing critical corporate or consumer data is the possible nightmare that looms over Fortune 500...

Software vulnerability detection has seen substantial advancements in integrating deep learning models, which have shown high accuracy in identifying potential vulnerabilities within software. These...

The proliferation of large language models (LLMs) across critical domains has highlighted the urgent need for frameworks to safeguard data privacy without sacrificing computational...

Despite the significant strides in large language models (LLMs) such as ChatGPT, Llama2, Vicuna, and Gemini, they grapple with safety issues. This paper introduces...

Cyberattackers use many methods to trick consumers into visiting malicious websites or giving over private information. One of the most popular forms of cyberattack...

The era of centralized databases has passed with greater trust in the online world. Many events have shown that the centralization of management and...

Cyber security and information security are commonly used interchangeably, but they differ. Not every piece of data can be information. Data, when given with...

In today's world, everyone is benefitting from internet connectivity. Just a click is enough to know about the whereabouts of a distant friend or...

The industry's standard procedure is to use the firmware update's current version number in conjunction with a public database of threats and vulnerabilities to...