search
AI security research
Trends
- 1Backdoor found in Claude Code extension on Cursor MarketplaceβBackdoor in Claude Code for VS Code extension on the Cursor Marketplace
A security researcher reports that the Claude Code VS Code extension listed on the Cursor Marketplace was compromised and contained a backdoor, putting developers who installed it at risk. The finding is drawing attention across the developer and security community, with discussion focused on how supply-chain attacks can slip through third-party extension marketplaces and what verification steps users should take.
- 2
The FBI arrested a Saudi engineer at OpenAI over a covert recording device, according to the Justice Department. The story lands amid a rough stretch for the company: it dismissed safety researchers, a departing employee warned AI firms are 'playing with fire', and it drew fresh criticism from mathematicians over claims of research breakthroughs. Commenters are weighing whether the arrest signals deeper internal security problems.
- 3Bruce Schneier warns AIs are hacking the legal systemβ# AIs # Hacking the # Law - https://www. schneier.com/blog/archives/202 6/11/ais-hacking-the-law.html "AIs finding vulne
Security researcher Bruce Schneier argues that artificial intelligence systems are now finding and exploiting vulnerabilities in the legal system the way hackers probe software. In a new blog post, he describes AIs uncovering loopholes and exploits in laws and regulations, raising concerns about how legal frameworks can be gamed at scale and what that means for lawmakers and regulators.
- 4Dutch researchers face record-low odds of winning grants as AI growsβResearchers face record-low chances of getting a grant partly due to rise of AI
Researchers in the Netherlands now have a record-low chance of securing research funding, according to Dutch news reporting. The article says the rise of artificial intelligence is one factor driving down success rates, as competition intensifies and more applications flood funding bodies. Scientists are voicing concern that fewer projects will get off the ground, with AI-related research drawing applicants away from other fields and squeezing overall grant budgets.
- 5Nebraska to lead cybersecurity for US AI-driven science networkβΌUniversity to lead cybersecurity for national Fabric for AI-Driven Science project
The University of Nebraska will lead the cybersecurity work for the national Fabric for AI-Driven Science project, a US research infrastructure initiative connecting institutions to support artificial intelligence-powered scientific research. The role puts the university in charge of protecting the network that researchers across the country will use for data-intensive, AI-driven science.
- 6patch2vuln tool updated with web interface and CVE supportβMany new changes for patch2vuln is a tool that turns a git-format patch into a structured draft vulnerability advisory u
patch2vuln, a tool that converts git-format patches into structured draft vulnerability advisories using a locally hosted Ollama model, has received a major update. The new version adds a web interface and API, along with significant changes to export format handling, including support for the CVE format. Security researchers are highlighting the release as a useful way to automate vulnerability advisory drafting while keeping AI processing local.
- 7
Anthropic's AI model has been credited with finding a vulnerability in the Linux kernel, tracked as CVE-2026-98432, which could allow attackers to gain remote root access to affected systems. Security outlet CSO Online reported the discovery, and the bug is drawing attention across searches in the UK, Germany, Russia and Japan as researchers assess its severity.
- 8AI Is Finding Software Vulnerabilities Faster Than Defenders Can RespondβΌAI can find serious vulnerabilities at unprecedented speed. Can defenders keep up?
IBM warns that artificial intelligence can now uncover serious software vulnerabilities at unprecedented speed, raising questions about whether security teams can keep pace. The concern is that the same tools accelerating discovery for researchers could be exploited by attackers, forcing defenders to rethink patching cycles and automated defenses.
- 9Anthropic says Claude can find real-world software vulnerabilitiesβDiscovering and exploiting vulnerabilities in real-world software with Claude
Anthropic has published research describing how its Claude AI model can discover and exploit security vulnerabilities in real-world software. The company outlines demonstrations of the model identifying flaws that could affect actual applications, a development with significant implications for both defensive security work and offensive capabilities. Security researchers are discussing what the findings mean for the future of automated vulnerability discovery, bug bounty work, and the risks of AI tools being used maliciously.
- 10Anthropic says Claude can find vulnerabilities in critical softwareβDiscovering and exploiting vulnerabilities in critical software with Claude
Anthropic has published research describing how its Claude AI model can be used to discover and exploit vulnerabilities in critical software. The writeup outlines the model's role in identifying security flaws that could affect widely used systems. Security researchers are discussing the implications of AI-assisted vulnerability discovery, including both its potential for defensive work and the risks of offensive use.
- 11OpenAI Pipeline Generates 948 Linux Root Escalation ExploitsβΌOpenAI AI Pipeline Generates 948 Linux Root Escalation Exploits in 30-Hour Run
OpenAI ran an automated AI pipeline for 30 hours that produced 948 Linux root escalation exploits, according to a report by The Hacker News. The scale of autonomously generated privilege-escalation code is drawing attention from security researchers concerned about offensive AI capabilities, and from defenders weighing how such tooling could be used to find and patch real vulnerabilities.
- 12
Security researchers report that Claude AI was able to exploit a Linux kernel vulnerability to execute root commands without any user interaction. The finding has raised alarm in the cybersecurity community, as it demonstrates how AI agents could be leveraged to autonomously exploit system flaws, prompting renewed debate about the safety risks of granting AI tools deep system access.
- 13Cybersecurity startup Cathedral researchers breached Anthropic's systemsβResearchers at cybersecurity startup Cathedral got inside Anthropic
Researchers at cybersecurity startup Cathedral reportedly gained access inside Anthropic, according to a Washington Post report. The disclosure raises questions about security practices at one of the leading AI companies and how a small startup managed to penetrate its defences. Details about the method and Anthropic's response remain limited.
- 14AI hacking vulnerabilities found in German legal systemβΌAI hacking vulnerabilities - in the German legal system: https://www. schneier.com/blog/archives/202 6/11/ais-hacking-th
Security expert Bruce Schneier has published an analysis of AI hacking vulnerabilities within the German legal system, examining how artificial intelligence tools could exploit weaknesses in laws and legal procedures. The piece is being shared in cybersecurity and legal circles, tagged under exploit and vulnerability discussions alongside CVE references.
- 15Chainstack discloses production breach tied to ARTEX attack campaignβΌChainstack Discloses Production Breach Linked to Global AI-Powered ARTEX Campaign Chainstack suffered a production breac
Blockchain infrastructure provider Chainstack has disclosed a breach of its production systems, which it links to a global AI-powered attack campaign dubbed ARTEX. The incident exposed customer account details and password hashes. The company says it has stopped the breach, but affected customers are being urged to change credentials, and security researchers are discussing the growing use of AI in such campaigns.
- 16Researchers warn of agentic backdoors planted via pretraining poisoningβΌAgentic backdoors via pretraining poisoning
A new research paper describes a method for planting hidden backdoors in AI models during pretraining, which then activate when the model is used in agentic settings. The work raises concerns that malicious actors could compromise models long before deployment, making the flaws hard to detect through standard safety testing.
- 17How Architecture Shapes User Behaviour OnlineβHow architecture shapes user behaviour # negativepid # digitalInvestigations # OSINT # cybersecurity # AI # tech # onlin
A blog post titled 'How architecture shapes user behaviour' is being shared, tagged with topics including OSINT, digital investigations, cybersecurity, AI, cyberpsychology and robotics. The post discusses how digital and physical design influences the way people act, a theme of interest to online investigators and security researchers. No further details from the post are available.
- 18
A new writeup describes how security researchers were able to take over ChatGPT user accounts, walking through the method and its impact. The disclosure is drawing wide attention among developers and security professionals, with discussion focused on how such an account takeover was possible and what it says about the security of widely used AI services.
- 19New guide details building a malware analysis lab with REMnuxβMy new guide explains how to build a malware analysis lab with REMnux. A Windows VM and REMnux share an isolated network
Security researcher Lenny Zeltser has published a guide on setting up a malware analysis lab using REMnux. The setup pairs a Windows virtual machine with REMnux on an isolated network, and includes an AI agent running on REMnux whose sole allowlisted external destination is the language model. The guide covers how analysts can safely examine malicious software without exposing their systems.
- 20Debate grows over true open source in AIβΌWhy we need actual Open Source # AI . Not just open weights...open source, open data, open eval, open alignment. # ML #
Arguments are resurfacing over what qualifies as genuinely open source artificial intelligence. Commenters argue that releasing model weights alone falls short, and that true openness requires open source code, open training data, open evaluation methods and open alignment processes. The discussion, framed through a machine learning security lens, is drawing attention among developers and security researchers weighing transparency against the industry's prevailing practice of open-weights releases.
- 21AI-powered Windows malware uses Claude to rank stolen victim dataβΌNew AI-driven Windows malware feeds stolen data to Claude to rank victimsβ assets by value, then deletes itself β payloa
Cisco Talos researchers have detailed VOLTA, a Windows infostealer that sends stolen victim data to Claude so the AI ranks the victim's assets by value, building a custom target assessment. The malware can switch to another model if one refuses the request, or skip the AI step entirely, and deletes itself after finishing. Security observers are highlighting the case as a notable example of criminals abusing mainstream AI tools in malware operations.
- 22UK school estate hit by 3AM ransomware in suspected AI-assisted attackβ# UK # school estate hit by # 3AM # ransomware in suspected AI-assisted attack https://www. bleepingcomputer.com/news/se
A UK multi-academy school estate has been struck by ransomware belonging to the 3AM group, in what security researchers suspect may be one of the first AI-assisted cyberattacks on education infrastructure. Details remain limited, but the incident is drawing attention for its potential use of artificial intelligence in the attack and its impact on schools' operations and data.
- 23UK school estate hit by suspected AI-assisted ransomware attackβΌReward: You've received a Ledger of Negative Assets. Current balance: 37 encrypted systems, two backdoor admins, and one
A ransomware attack struck a UK school estate at 3am, in what security researchers suspect was an AI-assisted intrusion. Commentators in the infosec community are circulating reports that dozens of encrypted systems were left behind, along with two backdoor administrator accounts, warning of a difficult recovery ahead. The case is fueling wider debate about attackers using AI tools to plan and execute intrusions, and about the growing exposure of educational institutions to ransomware gangs.
- 24UC San Diego's HalΔ±cΔ±oΔlu School Expands AI Computing InfrastructureβΌUC San Diegoβs HalΔ±cΔ±oΔlu School Expands AI Computing Infrastructure to Address Critical Shortages
UC San Diego's HalΔ±cΔ±oΔlu Data Science Institute, part of the university's School of Computing, Information and Data Sciences, is expanding its AI computing infrastructure to address critical shortages in computing capacity. The announcement, reported by UC San Diego Today, signals the school's effort to keep pace with surging demand for artificial intelligence research, teaching and computing resources.
- 25Criminals hijack corporate AI accounts, racking up $1 million billsβCriminals are hijacking corporate AI model access, leaving some victims with bills near $1 million Google says criminals
Google researchers report that criminals are using stolen cloud credentials to run expensive AI models on compromised corporate accounts. Some victims have faced bills approaching $1 million, with more than $100,000 in charges added in a single day. The attackers are reportedly reselling access to the hijacked AI capacity on underground markets.
- 26Kaspersky uncovers CyberStrikeAI framework weaponizing AI attacksβΌKaspersky uncovered CyberStrikeAI framework attacks using AI to automate post-exploitation and compromise domains across
Kaspersky researchers have uncovered attacks carried out with CyberStrikeAI, a framework that uses artificial intelligence to automate post-exploitation tasks after breaching a network. The campaign has been linked to compromises of domains in education-sector networks, including activity targeting Active Directory environments. The finding highlights how attackers are adopting AI tooling to speed up and scale intrusion operations, drawing attention from the cybersecurity community.
- 27Lenovo 'encrypted' USB drive revealed as unencrypted SD cardβΌLenovo's 'encrypted' fingerprint USB drive is just an unencrypted SD card in a fancy case β AI-assisted follow-up finds
Researchers report that Lenovo's ThinkPlus FU100 fingerprint USB drive, marketed as encrypted, actually contains a plain unencrypted SD card inside its case. An AI-assisted follow-up review allegedly uncovered backdoors and hardcoded encryption keys in other Lenovo 'secure' USB products as well, raising questions about the company's security claims and prompting discussion among security researchers online.
- 28Attackers deploy LLM-driven open-source cyberattack frameworkβNew. "This is an open-source framework in which an LLM orchestrator on the command-and-control server plans actions, sel
Security researchers are flagging CyberStrikeAI, an open-source framework in which a large language model orchestrator running on a command-and-control server plans actions, selects tools and generates commands, while an implant executes them on compromised hosts. Posts on an infosec forum report attackers using it to progress from reconnaissance to later attack stages, raising concern about AI-autonomous offensive tooling becoming freely available to less-skilled actors.
- 2928 Zero-Day Vulnerabilities Found in Under Two DaysβΌ28 Zero-Day Findings in Less Than Two Days: What AI-Speed Discovery Means for Robotics
Researchers reported uncovering 28 zero-day vulnerabilities in less than two days, crediting AI-assisted discovery for the speed of the findings. The result highlights how automated approaches could transform security testing for robotics, where software flaws in connected machines carry physical risks. Security experts are weighing what rapid, scalable vulnerability discovery means for defenders racing to patch industrial and consumer robots.
- 30Microsoft Ships 1,461 Patches for AI-Found BugsβΌAI-Found Bugs: Microsoft Ships 1,461 Patches [2026]
Microsoft has released 1,461 patches addressing bugs discovered with the help of artificial intelligence, according to reports dated 2026. The unusually large patch batch suggests AI-driven bug finding is scaling up security fixes across Microsoft software. Details on which products are affected and how many flaws were found by AI versus human researchers have not been confirmed.
- 31Open-source AI hacking tools linked to real cyberattacksβΌAs open-source artificial intelligence (AI) hacking tools have been used for actual cyberattacks, ex..
Open-source artificial intelligence hacking tools have reportedly been deployed in actual cyberattacks, prompting concern among security experts that freely available AI-driven offensive software is lowering the barrier for attackers. The development raises questions about how such tools are shared and controlled, and analysts are calling for closer oversight of open-source security research as it increasingly spills into criminal use.
- 32Carnegie Mellon Names Jason Brown as New SEI Directorβπππ«π§ππ π’π πππ₯π₯π¨π§ πππ¦ππ¬ πππ¬π¨π§ ππ«π¨π°π§ ππ¬ πππ° πππ ππ’π«ππππ¨π« Full analysis: https:// trendxinsights.com/press-relea ses/cmu-nam
Carnegie Mellon University has named Jason Brown as the new director of its Software Engineering Institute, according to an announcement. The appointment places Brown at the helm of the federally funded research and development center known for work on software engineering, cybersecurity and artificial intelligence. The news is being shared alongside commentary linking the move to growing institutional focus on AI and security research.
- 33Token thieves run up $100K+ daily in illicit LLM chargesβToken thieves costing companies $100K+ per day in illicit LLM charges
Attackers are stealing access to large language model services and burning through companies' budgets, with some firms reporting losses exceeding $100,000 per day in illicit compute charges. The thefts exploit leaked or stolen API credentials, allowing criminals to use paid AI capacity for free while victims foot the bill. Security coverage highlights the growing financial cost of unprotected AI infrastructure as adoption accelerates.
- 34Researchers uncover malware campaign targeting Hugging FaceβPluto Research exposes Hugging Face malware. This Hugging Face malware campaign drops cryptominers and backdoors on self
Security researchers at Pluto Research say they have exposed a malware campaign distributed through Hugging Face, the popular AI model-sharing platform. The campaign reportedly deploys cryptominers and backdoors onto self-hosted AI servers, posing a supply chain risk for organisations running their own AI infrastructure.
- 35
A report claims Microsoft released 1,461 security patches in 2026, describing the volume as an 'AI bug flood' and suggesting artificial intelligence code is generating an unprecedented number of fixes. Beyond the headline, no independent confirmation of the patch total, the affected products, or the causes has emerged, and security researchers have not yet publicly weighed in on the claim.
- 36New VOLTA Infostealer Uses AI to Rank Victims by WealthβΌNew VOLTA Infostealer Uses Claude, Gemini, and Kimi AI to Rank Victims by Financial Value
Cybersecurity researchers report a new infostealer malware called VOLTA that reportedly uses large language models including Claude, Gemini, and Kimi to analyse stolen data and rank victims by their financial value, allowing attackers to prioritise high-worth targets. The story is spreading through security news coverage, drawing attention to the growing use of commercial AI tools in criminal malware operations.
- 37New VOLTA Infostealer Uses AI to Rank Victims by Financial ValueβNew VOLTA Infostealer Uses Claude, Gemini, and Kimi AI to Rank Victims by Financial Value https:// thehackernews.com/202
Security researchers report a new infostealer malware called VOLTA that incorporates AI models, including Claude, Gemini, and Kimi, to assess and rank victims by their potential financial value. The malware reportedly uses these models to prioritize high-worth targets, marking a notable step in the criminal abuse of commercial AI tools. Coverage is spreading across cybersecurity communities, with commentators highlighting the implications of generative AI being repurposed for data theft operations.
- 38Google researchers warn of AI-driven agentic hacking frameworksβ"Security teams, developers, and system administrators will need to fundamentally rethink defense strategies to counter
Google Cloud threat intelligence researchers have published an analysis of Artex and related agentic offensive frameworks, in which AI agents can autonomously carry out cyberattacks at scale. The report argues that security teams, developers, and system administrators must fundamentally rethink defense strategies to counter scalable, autonomous offensive software.
- 39Microsoft details AI use by North Korean hackersβMicrosoft reveals how North Korean threat actors use tools like ARTEX. See how AI-enabled North Korean threat actors spe
Microsoft has published new threat intelligence describing how North Korean cyber actors are using AI-enabled tools, including one referred to as ARTEX, to accelerate their operations. The report outlines how automation and AI are speeding up reconnaissance and attack workflows, prompting discussion among security researchers about state-backed groups adopting emerging technology.
- 40VOLTA infostealer uses AI to pick its victimsβΌAppraisal by AI: Model-driven victim prioritization in the VOLTA infostealer
Cisco Talos researchers report that the VOLTA infostealer, a malware strain that harvests saved passwords and other credentials from infected computers, now uses an AI model to decide which victims are worth attacking. The malware scores potential targets before deploying its full payload, a sign that cybercriminals are adopting machine learning to make their operations more selective and efficient.
Repos
- affaan-m/ECC The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development f
- NVIDIA/SkillSpector Security scanner for AI agent skills. Detect vulnerabilities, malicious patterns, security risks, prompt injection, data