artificial intelligence | Breaking Cybersecurity News | The Hacker News

xAI's Grok Build coding CLI was uploading entire Git repositories, full commit history and all, to a Google Cloud Storage bucket run by xAI, not just the files a coding task needed. A researcher publishing as cereblab , testing version 0.2.93 , captured one of those uploads, cloned the git bundle out of the intercepted request, and pulled back a file the agent had been told in plain terms not to open. The upload rode a separate channel from the model itself, and the byte split is hard to argue with. On a 12 GB repo of files the model never read, model-turn traffic to /v1/responses came to about 192 KB while the storage channel to /v1/storage moved 5.10 GiB, a roughly 27,800x gap between what the model needed and what left the machine. That storage upload ran as 73 chunks of about 75 MB, every one returning HTTP 200, and across the researcher's size sweep the volume tracked total repo size. The destination bucket, grok-code-session-traces , is named in the binary and ...

Somewhere right now, a security tool is quietly finding bugs faster than any human can fix them. That's supposed to be the good news. The catch is that the attackers have the same tools, pointed the other way, and they don't file tickets. That's the shape of this week. Trusted code turns on the people who installed it. Old bugs from last year are still landing because the fix sat in a queue too long. Fake installers, poisoned packages, systems left facing the open internet, and helpful little AI assistants running instructions that were never yours. The gap between "patch exists" and "already exploited" keeps shrinking, and nobody's closing it. None of it is exotic. That's what wears you down. Same ordinary mistakes, just happening faster than we can keep up. Here's the full mess, top to bottom. ⚡ Threat of the Week Progress Tells ShareFile Customers to Shut Down Storage Zone Controllers — Progress urged customers to shut down Win...

Give an AI assistant a memory and access to your inbox, and you hand an attacker a way to rewrite what it thinks it knows about you. A single email can trick that agent into saving a false "fact" about the user, hide the change, and quietly steer its answers in later sessions. When it works, the person reads an ordinary-looking reply and never learns their assistant was tampered with. The researchers named the attack  stealth memory injection  and built a tool that writes the emails automatically. The paper, "When Claws Remember but Do Not Tell,"  landed on arXiv on 6 July 2026 . First, what these assistants do A personal agent is an AI assistant that sticks around. Instead of forgetting everything when a chat ends, it keeps notes about you in files: your preferences, your contacts, and what you asked it to do. It reads those notes at the start of every new session, which is why it feels like it knows you. Many of these agents can also act for you, readin...

A new phishing-as-a-service (PhaaS) operation called Forg365 is using a combination of device code phishing , adversary-in-the-middle (AitM) tactics, antibot evasion, artificial intelligence (AI)-assisted lure creation, and post-compromise mailbox operations targeting Microsoft 365 accounts. Distributed via Telegram and costing $400 a month (or $3,800 per year), attack chains leverage phishing lures that make use of legitimate email delivery infrastructure, such as Amazon Simple Email Service (Amazon SES) and Twilio SendGrid, to imitate a redirection chain that blends into regular email traffic before it ends in Forg365-controlled domains. "The panel exposes a mature operator workflow: accounts, links, invitations, OAuth app configuration, redirect links, SVG generation, campaign sending, SMTP profiles, SMTP rotation, AI email generation, token vaulting, account intelligence, keyword alerts, viewer links, and browser-extension support," ZeroBEC said . The email securi...

Meta has filed a patent application for an AI that listens to your voice throughout the day, works out how it thinks you are feeling from the way you sound, and keeps a timestamped log of every read. Each read gets pinned to the moment it happened: the time, your location, what you were doing, even how you were using your phone. Some versions in the filing would listen all day; others would check in only at set times. None of these ships in a product today, and Meta has not announced one; a filing like this stakes a claim on an idea long before anyone commits to building it. The application,  US 2026/0182881 , was filed by Meta Platforms in December 2025 and published on July 2. It names a single inventor, Lachlan Dunn , and traces back to a provisional filing from December 2024. The patent-analysis site  Patentlyze  flagged the filing first. Its title pairs two ideas, emotional state analysis and real-time fitness coaching. The claims show the first is the ...

A few days ago, I was sitting with the CISO of a Fortune 50 company, walking through how his security team was thinking about AI agents in the SOC. Smart team. Serious program. They had already connected Claude to a few detection tools and were seeing real value in specific investigations. But as we mapped out the broader architecture, something kept nagging at me. The design they were building was going to work beautifully for a tiny percentage of alerts that genuinely needed deep human judgment. It was going to completely ignore the rest. On the flight home, I picked up a book I had not touched in a few years. Daniel Kahneman's Thinking, Fast and Slow. Kahneman is one of the rare people who genuinely changed how we understand human decision-making. He spent his career as a psychologist studying how people actually think, as opposed to how economists assumed they did. In 2002, he won the Nobel Prize in Economics, which tells you something about how far his work traveled beyond ...

Cybersecurity researchers have flagged an intrusion in which an unknown threat actor leveraged a vibe-coded PowerShell script for Active Directory (AD) enumeration. "The script looked for the Domain Controller (DC) and mapped users, computers, and domains, before creating a directory and exporting out a number of files, and finally creating AD_Report.html to measure the success of the enumeration attempt," Huntress researchers Jevon Ang and Dray Agha said . The attack chain involved the threat actor establishing Remote Desktop Protocol (RDP) access onto a domain-joined Windows Server with a set of pre-compromised credentials, followed by staging the tools in the "C:\ProgramData\" folder. The incident took place in early June 2026. This included an artificial intelligence (AI)-generated payload to map the Active Directory environment. The assessment is based on various telltale signs, such as the prompt iteration title, placeholder strings, over-engineered cod...

An attacker running a live Microsoft 365 phishing operation left a Python web server listening on a public port with directory listing switched on. The command that did it:  python3 -m http.server 8080 , was still sitting in the readable  .bash_history . From that one lapse, French security firm  Lexfo  lifted the operator's entire toolkit and pivoted through it to two more phishing operators, three campaigns in all. Each ran a custom fork of the open-source Evilginx proxy , cloned from public GitHub. The largest of the three had been running for more than a year, its victims overwhelmingly corporate mailboxes. The three got past MFA in two mechanically different ways, one by proxying the live login , one by abusing a legitimate Microsoft sign-in flow. The two need different defenses, which is the part that matters most if you run Microsoft 365. Directory listing on a working attack server is close to a full confession. The listing exposed phishing conf...

Most security mess starts as admin work. A link gets clicked. A tool gets trusted. A bucket name gets reused. A setting stays loose because nobody wants to touch it. This week is full of that kind of damage. Not loud. Not clever. Just small gaps doing big jobs. The worst part is how normal it all looks until the bill arrives. The full ThreatsDay list is below. Global fraud bust Global Operation Leads to ~6K Arrests A global anti-fraud operation involving 97 countries and territories has resulted in the arrest of 5,811 individuals and the interception of $293 million in illicit assets as part of an operation codenamed First Light 2026 that took place between January 15 and April 30, 2026, to tackle social engineering scams and associated money laundering activities. "Over 142,000 victims globally were identified during Operation First Light 2026, highlighting the extent to which social engineering scams and fraud have escalated ...

Meta has announced that its new artificial intelligence (AI) model Muse Image lets people use public Instagram posts and reels to generate AI content, and it's enabled by default. "You can also @-mention Instagram accounts in the Meta AI app to bring specific Instagram profiles right into your images," the social media giant said in a post. "Whether you want to design a custom event invitation, mock up a collaborative creative concept, or generate a personalized graphic, tagging a username lets Meta AI use public photos to build a visual that's ready to post" Muse Image is Meta's inaugural image-focused AI model from its Superintelligence Labs, which the company said uses advanced reasoning to better understand complex prompts and blend multiple photos into high-quality creations for sharing across its platforms and elsewhere. It's also being embedded into WhatsApp and Instagram to facilitate AI-powered effects for Instagram Stories and ima...

An AI coding assistant that refuses to answer a dangerous request in its chat box can answer it anyway if the same request is broken into small, ordinary-looking steps inside a code editor. That is the finding of a  new study of GitHub Copilot  by researchers Abhishek Kumar and Carsten Maple. The models they tested through Copilot, Claude from Anthropic, and Gemini from Google, refused almost every harmful request when asked directly. Reframed as steps in a normal coding task, they produced the harmful answers in all 816 of the study's workflow runs. What makes this different from a typical jailbreak: no one asks for the harmful thing directly, and the model is not tricked into running someone else's code. It writes the banned content itself, as a side effect of a coding task it was told to improve. How it works The researchers call the method workflow-level jailbreak construction . Instead of a single blunt prompt, they asked Copilot to build an everyday piece of s...

A streaming box should not need a threat model. Neither should a username field, a demo repo, a reset flow, or a browser permission prompt. That is the irritating part this week: the risky pieces were ordinary. Home devices became a routing cover. Clean code pulled dirt from a dependency. Identity shortcuts aged badly. AI systems trusted the wrong instructions. Same soft spot throughout: trust placed one layer too early. Below is the full recap, since this is apparently what counted as a normal week. ⚡ Threat of the Week NetNut Residential Proxy Network Disrupted — Google, in collaboration with the U.S. Federal Bureau of Investigation (FBI), Lumen, and other partners, took action against the NetNut residential proxy network, also known as Popa, building upon its takedown of IPIDEA in January 2026. Google said it disabled Google accounts and associated Google services used by NetNut for malware command-and-control (C2) and updated Google Play Protect, in addition to disabling ...

Building a shortlist for an AI SOC evaluation can be tough. SIEM, SOAR, and pureplay AI SOC vendors are all saying the same thing. But behind the identical label sit very different products, from chat assistants bolted onto a legacy SIEM to agent platforms that run detection, triage, investigation, and response on their own data foundation. Whether a platform will materially change outcomes for your team matters more than what it is called. We can measure that in investigation time, false-positive volume, analyst hours returned, total cost of running your SOC and finally whether the architecture will hold up 2-3 years from now as the volume, speed and complexity of attacks keep increasing. What Is an AI SOC Platform? An AI SOC platform is a security operations platform where AI agents carry out the core work of the SOC (detection, triage, investigation, and response) by reasoning over correlated security data, under human oversight. It differs from bolt-on AI, which summarizes ...

Cybersecurity researchers have discovered a previously undocumented modular malware framework codenamed Avalon that's distributed by means of a multi-stage phishing chain capable of bypassing traditional security controls. Avalon combines credential collection, lateral movement, remote access, recovery disruption, and ransomware execution, bringing together diverse functions under one umbrella. The ransomware component has been internally named CrownX.  "The attack began with a spoofed legal document email directing recipients to a password protected archive on Proton Drive," Blackpoint Cyber researchers Nevan Beal and Sam Decker said . "Malicious content was embedded inside an ISO image rather than attached directly, reducing the likelihood of detection at the email layer." Should the email recipient interact with a document-themed Windows Shortcut ("Secure Document CA-283505.pdf.lnk") inside the mounted image, it triggers a staged malware s...

This week’s security news is mostly about weak spots. Browsers, bots, sandboxes, AI systems, and email flows all show the same problem in different ways. Everything looks normal until someone tests a small gap and finds a way through. This is not one big break. It is small permissions, weak checks, open systems, and normal tools doing things they were allowed to do. That same pattern runs through the stories below.

Identity lifecycle management was architected around a person with an employment record, a manager, and a departure date. AI agents have none of those. As autonomous principals proliferate across enterprise environments, the governance model built for humans develops structural blind spots that traditional IGA tools weren't designed to detect. This guide covers where that model breaks, what it fails to govern, and what extending it to agents actually requires. What Identity Lifecycle Management Was Designed to Handle To understand why identity lifecycle management breaks down around AI agents, you need to understand what it was built to do well and who it was built for. The entire architecture rests on a single foundational assumption: every identity maps to a human being whose organizational status changes through documented, HR-driven events. The identity lifecycle management process governs access from an identity's first provisioning event through every modificatio...

Security firm Sysdig says it has found what it believes is the first ransomware attack run from start to finish by an AI agent. Its Threat Research Team calls the operator JADEPUFFER and says a large language model handled the whole job: breaking in, stealing credentials, moving deeper into the network, then encrypting and wiping a company's production database. Ransomware has always needed a skilled person somewhere in the loop, either at the keyboard or writing the script the malware follows. If a model can chain those steps on its own, the skill needed to run an attack drops to whatever it costs to rent an AI agent. The way in was an old, already-patched bug. JADEPUFFER exploited  CVE-2025-3248 , a missing-authentication flaw in  Langflow , an open-source tool for building AI apps and agent workflows. The flaw lets anyone who can reach the server run their own Python code on it, no login needed. Langflow boxes are a tempting target because they often sit ...

Adobe has released patches for multiple maximum-severity security flaws impacting Adobe ColdFusion and Adobe Campaign Classic. The ColdFusion updates "resolves critical and important vulnerabilities that could lead to arbitrary code execution, privilege escalation, arbitrary file system read, and security feature bypass," Adobe said in an alert released Tuesday. The vulnerabilities are listed below - CVE-2026-48276, CVE-2026-48283 (CVSS scores: 10.0) - Unrestricted upload of file with dangerous type vulnerabilities that could lead to arbitrary code execution CVE-2026-48277, CVE-2026-48281, CVE-2026-48316 (CVSS scores: 10.0) - Improper input validation vulnerabilities that could lead to arbitrary code execution CVE-2026-48282 (CVSS score: 10.0) - A path traversal vulnerability that could lead to arbitrary code execution CVE-2026-48313 (CVSS score: 9.3) - A path traversal vulnerability that could lead to arbitrary file system read CVE-2026-48315 (CVS...

Organizations have never had greater awareness of cyber risk. Yet turning that awareness into operational resilience has never been more challenging. The 2026 Bitdefender Cybersecurity Assessment confirms this is the case, as this year's findings reveal a series of surprising contradictions. Here are a few examples, based on the independent survey of 1,200 IT and cybersecurity professionals across six countries. IT & security leaders believe they have sufficient visibility into employee AI usage, while many frontline practitioners disagree .  Security teams understand the importance of reducing the attack surface, yet they often lack the skills, resources, or strategy to do so.  AI dominates cybersecurity conversations, but in some cases, it is drawing attention away from more prevalent attack techniques already causing significant damage.  Although organizations say they recognize the importance of transparency after a breach, many professionals st...

Large language models keep inventing web addresses that do not exist. Attackers have started buying those made-up domains before anyone else can, then hosting phishing pages on them to catch traffic that AI tools point their way. Palo Alto Networks' Unit 42 calls the trick phantom squatting , and its new research shows it is already happening in the wild. The reason it matters is trust. Developers and AI assistants increasingly treat the links a model hands back as real. When a model invents a domain that does not exist yet, whoever registers it first inherits all of that misplaced trust, with no phishing email and no malicious ad required. To measure the problem, Unit 42 asked two AI models 685,339 questions about 913 well-known brands across technology, finance, healthcare, government, gambling, and other sectors. The models produced 2.1 million links. Threat intelligence already flagged 13,229 of them as outright malicious, meaning the AI was handing out known-ba...