AI Security
79 articles
AI security, LLM abuse, model governance, and defender AI coverage.
HIGHAI Agents Automate Exploitation of Obscure Vulnerabilities
AI agents now discover and exploit obscure vulnerabilities autonomously, while AI-generated code floods pipelines with flaws. Defenders must adapt to agent-scale threats.
HIGHChrome 148 Patches AI Site Isolation Bypass, Android Payment Flaw
CVE-2026-8568 (CVSS 3.1) lets attackers bypass Chrome Site Isolation via AI features after renderer compromise; CVE-2026-8566 (CVSS 4.3) targets Android Payments.
HIGHOpen WebUI Patches Three Flaws: XSS, SVG Injection, Auth Bypass
Open WebUI fixes CVE-2026-45314 (SVG XSS), CVE-2026-45303 (iframe script injection), and CVE-2026-44567 (pending role auth bypass) — all in self-hosted AI platform.
HIGHAI Hallucinations Exploit Human Trust in Critical Infrastructure
AI models produce confident but incorrect outputs that have led to misconfigured firewalls and pipeline valve errors, researchers warn.
INFORMATIONALAI Security Startup Funding Surpasses Acquisitions by $1B in 1Q26
Dark Reading reports AI security startup investments exceeded acquisition value by over $1 billion in 1Q26, signaling a widening 'valley of death' for maturing firms.
HIGHHackers Exploit PraisonAI Auth Bypass Hours After Disclosure
Sysdig detected CVE-2026-44338 exploitation attempts within 3 hours 44 minutes of public advisory — attackers probed /agents on exposed PraisonAI instances.
INFORMATIONALMythos AI Excels at Code Audits but Struggles With Exploit Validation
XBOW benchmarks show Anthropic's Mythos AI is potent for source code audits and reverse engineering, but inconsistent at exploit validation and prone to overstating findings.
HIGHOpenAI Breached in TanStack Supply Chain Attack
OpenAI says two employees' devices were compromised in the TeamPCP Mini Shai-Hulud campaign, forcing rotation of code-signing certificates across macOS, Windows, iOS, and Android.
CRITICALPwn2Own Berlin 2026: Researchers Earn $523K Hacking Windows 11, Edge
On day one of Pwn2Own Berlin 2026, researchers collected $523,000 for 24 zero-days, including a $175,000 Edge sandbox escape by Orange Tsai and three Windows 11 privilege...
HIGHAI-Driven Attacks Compromise Systems in 73 Seconds, Outpacing Patching
Picus Security analysis shows AI-powered attackers exploit CVEs in ~10 hours and breach systems in 73 seconds, while patching still takes 24 hours.
CRITICALCosyVoice gRPC Server Insecure Deserialization Flaw CVE-2026-31251
CVE-2026-31251: CosyVoice gRPC server deserializes untrusted models via torch.load() without weights_only=True, enabling RCE via crafted .pt files. No patch confirmed.
HIGHFake OpenAI Repo on Hugging Face Pushes Rust Infostealer
A typosquatted OpenAI repository reached #1 on Hugging Face with 244,000 downloads, delivering a Rust-based infostealer that targets browser credentials, crypto wallets, and VPN...
HIGHPraisonAI Flaw Lets Agents Execute Arbitrary Python Tools
CVE-2026-44339 (CVSS 8.6) in PraisonAI multi-agent framework lets agents resolve undeclared tool names against module globals, enabling arbitrary Python execution.
HIGHBraintrust Breach Exposes AI Provider API Keys, Urges Rotation
Braintrust disclosed a breach on May 4 where attackers accessed an AWS account, compromising AI provider API keys for firms like Box and Stripe. At least one customer affected.
Boost Security Raises $4M, Acquires SecureIQx and Korbit.ai
Boost Security raised $4M to expand its AI-native SDLC defense platform, acquiring SecureIQx for reachability analysis and Korbit.ai for code review.
CRITICALCritical Ollama Bug CVE-2026-7482 Exposes 300K Deployments
Cyera discloses CVE-2026-7482 (CVSS 9.3) — a heap out-of-bounds read in Ollama's GGUF model loader that leaks prompts, API keys, and secrets via three unauthenticated API calls.
Cisco Acquires Astrix Security for Non-Human Identity Protection
Cisco announced plans to acquire Astrix Security to address non-human identity risks in AI and machine workloads. The deal expands Cisco's identity security portfolio.
MEDIUMOpenAI Strengthens ChatGPT Login Security With New Controls
OpenAI rolls out Advanced Account Security for ChatGPT: mandatory passkeys, shorter sessions, and account recovery changes. Affects all users globally.
HIGHAI-Assisted Attacks Reshape Cyber Threat Landscape in 2026
Japan teen arrested for stealing 7M Kaikatsu Club records to buy Pokémon cards — one example of how AI tools are lowering the barrier for cybercrime in 2026, per The Hacker News…
HIGHAI Agents Wreck Production Databases Due to Poor Access Controls
Dark Reading reports AI agents are deleting production databases because organizations deploy agent integrations without proper security testing or access controls.
MEDIUMUK Cyber Agency Warns AI Will Trigger 'Patch Wave' of Urgent Fixes
NCSC warns organizations to brace for a surge of urgent patches as AI accelerates vulnerability discovery, raising exploitation risk. No specific CVEs cited.
HIGHAnthropic Launches Claude Security for AI-Driven Exploit Defense
Anthropic released Claude Security, a defensive AI suite to counter autonomous exploit tools like Mythos that weaponize zero-days in minutes. Targets enterprise SOCs.
HIGHAI Browser Extensions Steal Emails, Passwords via Prompt Injection
Unit 42 finds 30+ malicious AI browser extensions exfiltrating email content, credentials, and API keys via prompt injection and DOM scraping. Affects Chrome, Edge users.
HIGHBluekit Phishing Service Offers AI Assistant, 40 Templates
A new phishing-as-a-service platform called Bluekit provides over 40 templates targeting banks, social media, and email providers, plus an AI assistant for drafting lures.
HIGHZero-Window Era: NDR Playbooks for Post-Mythos Exploits
Claude Mythos and Project Glasswing shrink exploit windows to near-zero. The Hacker News details NDR playbooks to contain AI-driven attacks before patching is possible.
CRITICALLiteLLM CVE-2026-42208 Pre-Auth SQLi Exploited in Attacks
Attackers exploit CVE-2026-42208, a critical pre-authentication SQL injection in LiteLLM LLM gateway, to steal API keys and model data. CVSS 9.8. No patch yet.
HIGHBlueNoroff Fakes Zoom Calls to Lure Crypto Execs
BlueNoroff uses stolen video, AI avatars, and fake Zoom invites to turn crypto executives into attack lures.
HIGHAI Assistants Reshape Security Priorities for Enterprises
Autonomous AI agents with file and service access are forcing organizations to rethink identity controls, data boundaries, and monitoring — Krebs reports on shifting attack…
HIGHMythos AI Finds Bugs Faster Than Teams Can Patch
Anthropic's Claude Mythos Preview identifies vulnerabilities at scale since April 7, but organizations lack the triage and patching capacity to keep pace, researchers warn.
MEDIUMGoogle: AI Prompt Injection Attacks Rising, Still Low-Sophistication
Google reports a rise in malicious AI prompt injection attacks, but most remain low-sophistication and harmless. Indirect injection attempts target LLM-integrated apps.
MEDIUMESET: March 2026 Cyber Threats Show Resilience Gaps
ESET's Tony Anscombe warns that March 2026 attacks — including ransomware, supply chain compromises, and AI-driven phishing — reveal systemic gaps in organizational…
MEDIUMRecorded Future: Malicious Infrastructure Evolves with AI-Driven
Insikt Group's 2025 Malicious Infrastructure Report tracks shifts in Cobalt Strike, Vidar infostealers, and AI-driven hosting tactics to inform defender strategies for 2026.
INFORMATIONALElastic Security Backs UK MoD Defence Cyber Marvel 2026 Exercise
Elastic Security Labs deployed AI-driven detection pipelines for the UK Ministry of Defence's Defence Cyber Marvel 2026 exercise, processing 1.2TB of telemetry across 50 simulated…
HIGHAI Agent Authority Gap Creates New Enterprise Security Blind Spots
The Hacker News reports AI agents create a structural security gap: delegated actors lack continuous oversight, enabling lateral movement and privilege escalation without human…
HIGHAI-Powered Phishing Surges as Attackers Personalize Lures at Scale
Enterprises report a sharp rise in AI-generated phishing campaigns that craft personalized lures at scale, moving from broad sprays to 1-to-1 targeting in the last six months.
Copperhelm Raises $7M for Agentic Cloud Security Platform
Copperhelm, an Israel-based startup founded by ex-RSA and McAfee engineers, raised $7 million in seed funding for an agentic AI platform that autonomously hunts cloud…
HIGHLMDeploy SSRF Flaw CVE-2026-33626 Exploited 13 Hours After Disclosure
CVE-2026-33626 (CVSS 7.5) in LMDeploy, an open-source LLM toolkit, was exploited in the wild within 13 hours of public disclosure, enabling SSRF attacks to access sensitive…
HIGHShadow AI and SaaS Expand Enterprise Attack Surface
Forgotten integrations, shadow IT, and unmanaged SaaS agents create new attack vectors. Dark Reading reports attackers exploit these gaps without sophisticated AI.
MEDIUMUS Vows Crackdown on Chinese Firms Exploiting American AI Models
Trump administration announces policy to penalize Chinese companies exploiting U.S. AI models via reverse engineering or unauthorized access, citing national security risks.
HIGHNorth Korean Hackers Steal $12 Million in Crypto via Trojanized
North Korean hackers siphoned over $12 million from crypto users in Q1 2026 using trojanized trading apps like CoinStats and TradingView AI Agent to steal recovery phrases and…
HIGHPalo Alto Networks Zealot AI Agent Autonomously Hacks Cloud Systems
Palo Alto Networks researchers built Zealot, a multi-agent AI penetration testing PoC that autonomously performs reconnaissance, exploitation, and data exfiltration on cloud…
HIGHAgentic AI Systems Introduce Novel Enterprise Security Risks
Recorded Future warns that autonomous 'agentic' AI systems, now being integrated into enterprise software, create new attack surfaces for prompt injection, data poisoning, and…
CRITICALCohere AI Terrarium Sandbox Flaw Allows Root Code Execution,
CVE-2026-5752 (CVSS 9.3) in Cohere AI's Terrarium sandbox enables root-level code execution and container escape via JavaScript prototype chain traversal.
HIGHFake TradingView AI Agent Site Drops Browser-Hijacking Malware
A malicious website impersonating a TradingView AI agent deploys malware that hands attackers full control of victims' browsers, enabling account theft and financial data…
BreachLock Named in Gartner Market Guide for Adversarial Exposure Validation
BreachLock is named a representative vendor in Gartner's 2026 Market Guide for Adversarial Exposure Validation, a category focused on AI-driven, continuous security testing.
HIGHFake Google Antigravity Installer Steals Accounts via Trojanized AI Tool
Malwarebytes reports a trojanized installer for Google's Antigravity AI tool steals browser cookies and account credentials within minutes, targeting users seeking the leaked software.
HIGHNGate Malware Trojanizes HandyPay App to Steal Brazilian NFC Data
NGate malware, using AI-generated code, has infected the legitimate HandyPay NFC app to steal payment card data and PINs from over 220,000 Android users in Brazil, according to ESET.
HIGHNGate Malware Uses AI to Evade Detection in Trojanized NFC Apps
NGate malware version 2.0, built with AI assistance, hides in a trojanized NFC payment app to steal SMS, contacts, and crypto wallet data from Android devices while evading security software.
CRITICALSGLang Vulnerability CVE-2026-5760 Enables Remote Code Execution via GGUF Files
CVE-2026-5760, a critical 9.8 CVSS flaw in the SGLang inference engine, allows attackers to execute arbitrary code by uploading malicious GGUF model files, compromising AI/ML serving deployments.
HIGHNorth Korean Operatives Use AI and Fake Identities to Infiltrate Companies via
North Korean operatives are using AI tools and forged documents to pass remote job interviews, according to Flare research. The tactic aims to place threat actors inside target companies for long-term espionage and network access.
HIGHVercel Breach Exposes Customer Credentials via Compromised AI Tool
Vercel confirms a breach exposing limited customer credentials after attackers compromised an employee's account via a third-party AI tool, Context.ai. The cloud platform is resetting passwords and API tokens for affected users.
HIGHAI-Powered Vulnerability Discovery Accelerates Exploit Timelines, Strains
Qualys warns that AI agents like Claude Mythos can cut vulnerability discovery time from months to hours, compressing the patch window and overwhelming security teams with a surge of new CVEs.
HIGHCloud Security Alliance Warns of AI Vulnerability Storm Post-Mythos
The Cloud Security Alliance warns that Anthropic's Claude Mythos model will trigger an 'AI vulnerability storm,' forcing CISOs to manage a 10x surge in code flaws and novel exploit techniques within 18 months.
HIGHFlorida Investigates ChatGPT Role in Campus Shooting Threat
Florida law enforcement is investigating how a student used ChatGPT to craft a threat of a campus shooting, part of a broader pattern where AI chatbots fail to block dangerous content.
HIGHOpenClaw AI Agent Poses Autonomous Threat via Package Masquerade
Qualys ETM detected the OpenClaw AI agent disguised as a routine package on a Windows Server, correlating endpoint, exposure, and identity telemetry to reveal an active, autonomous threat.
HIGHPushpaganda Scam Hijacks Google Discover with AI-Generated News
The Pushpaganda campaign used AI-generated news sites to poison Google Discover, pushing 10,000+ deceptive articles to trigger browser notifications that delivered scareware and ad fraud to millions of users.
MEDIUMBusiness Impersonation Fraud Evolves with AI-Powered Shopping Scams
Recorded Future details how threat actors exploit corporate identity verification gaps, pivoting from cashing stolen checks to orchestrating AI-powered shopping scams that impersonate legitimate businesses to steal goods.
HIGHAnthropic Restricts Access to AI Model Capable of Automated Vulnerability
Anthropic has restricted its Claude Mythos Preview AI to ~50 critical infrastructure vendors, citing its advanced ability to autonomously find and exploit software vulnerabilities, raising concerns about dual-use risks and offensive cyber capabilities.
INFORMATIONALAxonius Expands Asset Cloud with AI Remediation and OT Security
Axonius has updated its Asset Cloud platform with AI-powered remediation for exposures, added IoT/OT asset management, and introduced an asset trust standard to quantify security posture.
INFORMATIONALGitLab 18.11 Expands Agentic AI to Security Remediation and CI Pipelines
GitLab 18.11 integrates agentic AI across the software lifecycle, automating security fix generation and CI/CD pipeline configuration, aiming to address the 'AI paradox' of rapid code creation outpacing security and delivery.
INFORMATIONALAI SOC Tools Criticized for Automating Triage, Not Reducing Analyst Workload
A new analysis argues most AI-powered security operations center tools merely accelerate alert triage without reducing the underlying workload for analysts, failing to deliver on promises of true automation.
INFORMATIONALAnthropic Releases Claude Opus 4.7 with Automated Cybersecurity Safeguards
Anthropic releases Claude Opus 4.7, a frontier AI model with new automated safeguards designed to detect and halt potentially harmful cybersecurity tasks during long, unsupervised agentic workflows.
INFORMATIONALOpenAI Expands Access to GPT-5.4-Cyber for Defensive Security Tasks
OpenAI is expanding access to its GPT-5.4-Cyber model, a specialized AI for reverse engineering and malware analysis, following the reveal of Anthropic's offensive-capable 'Mythos' model. The move aims to lower barriers for legitimate security research.
HIGHPentera Report Warns of Critical Security Gaps in Agentic AI Architectures
Pentera's 2026 AI Security and Exposure Report finds that 100% of surveyed organizations have AI security gaps, with agentic AI architectures introducing novel risks like prompt injection and data exfiltration through deterministic workflows.
HIGHATHR Vishing Platform Automates Voice Phishing with AI Agents
The ATHR cybercrime platform automates voice phishing (vishing) attacks using AI-generated voice agents to impersonate trusted entities and harvest credentials, lowering the barrier for large-scale social engineering campaigns.
HIGHClickFix Phishing Campaign Masquerades as Claude AI Installer
A phishing campaign uses fake Claude AI installer lures and 'ClickFix' social engineering to trick users into granting remote access, enabling credential theft and financial fraud.
MEDIUMPushpaganda Campaign Exploits Google Discover to Hijack Browser Notifications
A threat operation dubbed Pushpaganda is abusing Google Discover with AI-generated clickbait to trick users into enabling malicious browser notifications, which then deliver phishing and scam content.
MEDIUMAsia's Digital Supply Chain Poses Distinct Security Challenges
Asia's interconnected digital ecosystems, divergent regulatory regimes, and rapid AI adoption are creating unique and complex security risks for regional and global supply chains, according to a new analysis.
MEDIUMPushpaganda Campaign Uses AI-Generated Clickbait to Hijack Browser Notifications
A campaign dubbed Pushpaganda uses AI-generated clickbait to trick users into enabling malicious browser notifications, delivering a persistent stream of scams and fake alerts directly to the desktop.
HIGHCSA Warns of AI-Driven 'Mythos' Era Collapsing Vulnerability-to-Exploit Timelines
The Cloud Security Alliance warns that AI models like Mythos are dramatically accelerating cyberattacks, collapsing the time between vulnerability discovery and weaponized exploit to near zero.
MEDIUMAI Chatbots as Political Advisors Raise Security and Transparency Concerns
A U.S. Senator's use of an AI chatbot for policy consultation highlights emerging risks in AI-assisted governance, including data privacy, model integrity, and accountability gaps.
HIGHFake Claude AI Website Delivers PlugX RAT via DLL Sideloading
A fraudulent website impersonating Anthropic's Claude AI distributes a self-deleting installer that deploys the PlugX remote access trojan via DLL sideloading.
HIGHAI Browser Extensions: The Unseen Threat Vector in Enterprise Networks
A new report from LayerX highlights the significant security risks posed by AI browser extensions, which are often overlooked in enterprise networks.
HIGHNorth Korean Lazarus Group Compromises OpenAI via Axios Supply Chain Attack
North Korea's Lazarus Group compromised OpenAI's internal systems via a supply chain attack on the Axios client library, using a stolen macOS code-signing certificate to sign malware.
HIGHFake Claude AI Site Delivers PlugX Malware in Trojanized Installer
A sophisticated phishing campaign uses a counterfeit Claude AI website to distribute a trojanized installer, deploying the remote access trojan PlugX to establish persistent backdoor access.
MEDIUMOpenAI Removes ChatGPT Study Mode, Raising Security and Transparency Concerns
OpenAI has removed the undocumented 'Study Mode' from ChatGPT, a feature that disabled web search and file uploads, highlighting concerns over silent feature changes and potential security implications for automated workflows.
MEDIUMOrange Business Integrates AI into Enterprise Voice, Raises Security Questions
Orange Business is embedding generative AI into its enterprise voice platforms, a move that expands the attack surface and introduces novel data security and privacy risks.
HIGHAI-Powered Threat Actor Breaches Mexican Government, Exposes Citizen Data
A sophisticated attacker leveraged AI tools like Claude and ChatGPT to breach nine Mexican government agencies, exfiltrating hundreds of millions of citizen records in a multi-month campaign.
HIGHGPT-5 Release: Security Implications for Enterprise Defenders
OpenAI's GPT-5 raises the bar for AI-assisted cyberattacks — spear-phishing at scale, automated exploit generation, and deepfake social engineering. Here's what security teams need to know and do.
Stay Updated
Get the latest cybersecurity news delivered to your inbox.