Search results for GPT have tokens | Breaking Cybersecurity News | The Hacker News

U.S. cybersecurity and intelligence agencies have accused China-based artificial intelligence (AI) companies of conducting "systematic extraction" of proprietary functionalities and capabilities of American frontier models through distillation attacks. The activity has been described as occurring at an industrial-scale and one that forms the "core" of their AI development strategy, according to a bulletin released by the National Security Agency (NSA), the Cybersecurity and Infrastructure Security Agency (CISA), and the Federal Bureau of Investigation (FBI). "While 'distillation' is recognized as a legitimate and useful technique in AI research, China-based AI companies are engaging in aggressive, malicious, and targeted distillation activities at an industrial scale that extract restricted proprietary functionalities and capabilities of U.S. frontier AI models," the authoring agencies said . The joint advisory noted that Chinese AI firms l...

OpenAI on Friday released three versions of GPT-5.6 , called Sol, Terra, and Luna , as a limited preview to a small number of companies as part of an ongoing engagement with the U.S. government. While Sol is the latest flagship model and the most powerful, Terra strikes a balance between efficiency and power, and Luna is fine-tuned for speed and affordability. "GPT‑5.6 Sol launches with our most robust safety stack to date. We strengthened protections for higher-risk activity, sensitive cyber requests, and repeated misuse, and spent multiple weeks finding weaknesses, pressure-testing our system, and hardening it against real-world attacks," OpenAI said . The model has also been touted as the "most capable model yet" for cybersecurity, making it much more suitable for vulnerability research and exploitation. On ExploitBench , GPT‑5.6 Sol is competitive with Anthropic Mythos Preview using only about one-third of the output tokens, OpenAI noted. The goal, ...

Hackers aren’t kicking down the door anymore. They just use the same tools we use every day — code packages, cloud accounts, email, chat, phones, and “trusted” partners — and turn them against us. One bad download can leak your keys. One weak vendor can expose many customers at once. One guest invite, one link on a phone, one bug in a common tool, and suddenly your mail, chats, repos, and servers are in play. Every story below is a reminder that your “safe” tools might be the real weak spot. ⚡ Threat of the Week Shai-Hulud Returns with More Aggression — The npm registry was targeted a second time by a self-replicating worm that went by the moniker "Sha1-Hulud: The Second Coming," affecting over 800 packages and 27,000 GitHub repositories. Like in the previous iteration, the main objective was to steal sensitive data like API keys, cloud credentials, and npm and GitHub authentication information, and facilitate deeper supply chain compromise in a worm-like fashion. Th...

Cybersecurity researchers have discovered a malicious Google Chrome extension that's designed to steal data associated with Meta Business Suite and Facebook Business Manager. The extension, named CL Suite by @CLMasters (ID: jkphinfhmfkckkcnifhjiplhfoiefffl), is marketed as a way to scrape Meta Business Suite data, remove verification pop-ups, and generate two-factor authentication (2FA) codes. The extension has 33 users as of writing. It was first uploaded to the Chrome Web Store on March 1, 2025. However, the browser add-on also exfiltrates TOTP codes for Facebook and Meta Business accounts, Business Manager contact lists, and analytics data to infrastructure controlled by the threat actor, Socket said. "The extension requests broad access to meta.com and facebook.com and claims in its privacy policy that 2FA secrets and Business Manager data remain local," security researcher Kirill Boychenko said . "In practice, the code transmits TOTP seeds and current one-t...

A lot of security problems still begin with someone doing a completely normal thing. Cloning a repo. Answering a call. Leaving a box exposed. Trusting the default. That pretty much covers the mood this week. Old bugs are back, supply chains are getting stranger, and some exploit paths are so short you wonder what was supposed to stop them in the first place. That’s only part of it. Here’s everything else that made the Monday recap. ⚡ Threat of the Week Anthropic's Model Attempts to Poison Open-Source Project — A new evaluation conducted by the U.K. AI Security Institute (AISI) found that AI models with access to the internet reached out into the real world to target individuals and organizations autonomously across 10 of the total of 122 runs. Of 19 such actions recorded, 17 originated from Anthropic's Mythos 5 and the remaining two involved OpenAI's GPT-5.6-Sol with cyber classifiers. In the most serious case, Anthropic's Claude Mythos 5 spent 34 hours trying...

AI agents promise to automate everything from financial reconciliations to incident response. Yet every time an AI agent spins up a workflow, it has to authenticate somewhere; often with a high-privilege API key, OAuth token, or service account that defenders can’t easily see. These “invisible” non-human identities (NHIs) now outnumber human accounts in most cloud environments, and they have become one of the ripest targets for attackers. Astrix’s Field CTO Jonathan Sander put it bluntly in a recent Hacker News webinar : “One dangerous habit we’ve had for a long time is trusting application logic to act as the guardrails. That doesn’t work when your AI agent is powered by LLMs that don’t stop and think when they’re about to do something wrong. They just do it.” Why AI Agents Redefine Identity Risk Autonomy changes everything: An AI agent can chain multiple API calls and modify data without a human in the loop. If the underlying credential is exposed or overprivileged, each addit...

Losing sleep over Generative-AI apps? You're not alone or wrong. According to the Astrix Security Research Group, mid size organizations already have, on average, 54 Generative-AI integrations to core systems like Slack, GitHub and Google Workspace and this number is only expected to grow. Continue reading to understand the potential risks and how to minimize them.  Book a Generative-AI Discovery session with Astrix Security's experts (free - no strings attached - agentless & zero friction) "Hey ChatGPT, review and optimize our source code"  "Hey Jasper.ai, generate a summary email of all our net new customers from this quarter"  "Hey Otter.ai, summarize our Zoom board meeting" In this era of financial turmoil, businesses and employees alike are constantly looking for tools to automate work processes and increase efficiency and productivity by connecting third party apps to core business systems such as Google workspace, Slack and GitHub ...

Google on Wednesday announced Gemini 3.8 Flash Cyber, which it described as its most capable cybersecurity model, and has made it available to a set of trusted defenders via a new initiative called the Fairwind Program . "The Fairwind Program gives high-priority defenders (like governments, healthcare providers, and telecommunications services) early access to advanced models that help them build better defenses, before new threats arrive," Google said . "So defenders have an early advantage, to help them protect vital infrastructure – which in turn protects people who rely on those systems." The tech giant said it's currently working with over 650 partners globally, including CrowdStrike, Datadog, Menlo Security, Palo Alto Networks, and Snowflake. The program is available to a group of Google Cloud customers, government agencies, and cybersecurity partners. The release of Gemini 3.8 Flash Cyber comes a little over a month after Google unveiled Gemini 3....

This week was a reminder that attackers do not always need big tricks. One small mistake, one old access path, one missed patch, and suddenly the door is open. The noise is not all noise, either. Forums are talking, researchers are finding easy cracks, and defenders have more cleanup waiting. Here’s the full Monday recap. ⚡ Threat of the Week New DirtyClone Linux Kernel Flaw Lets Local Users Gain Root via Cloned Packets — Cybersecurity researchers detailed a new variant of the Dirty Frag Linux kernel flaw. Called DirtyClone (aka CVE-2026-43503), it allows local users to gain root privileges via cloned packets. The exploit works successfully on Debian, Ubuntu, and Fedora systems with default namespace configurations. "Any local user on a server or device running a vulnerable kernel who holds or can acquire the CAP_NET_ADMIN capability (frequently obtainable via unprivileged user namespaces) [is exploitable]," JFrog said. "This poses the highest risk to multi-te...

Microsoft has disclosed details of a novel side-channel attack targeting remote language models that could enable a passive adversary with capabilities to observe network traffic to glean details about model conversation topics despite encryption protections under certain circumstances. This leakage of data exchanged between humans and streaming-mode language models could pose serious risks to the privacy of user and enterprise communications, the company noted. The attack has been codenamed Whisper Leak . "Cyber attackers in a position to observe the encrypted traffic (for example, a nation-state actor at the internet service provider layer, someone on the local network, or someone connected to the same Wi-Fi router) could use this cyber attack to infer if the user's prompt is on a specific topic," security researchers Jonathan Bar Or and Geoff McDonald, along with the Microsoft Defender Security Research Team, said . Put differently, the attack allows an attacker t...

A new attack campaign has targeted known Chrome browser extensions, leading to at least 35 extensions being compromised and exposing over 2.6 million users to data exposure and credential theft. The attack targeted publishers of browser extensions on the Chrome Web Store via a phishing campaign and used their access permissions to insert malicious code into legitimate extensions in order to steal cookies and user access tokens. The first company to shed light the campaign was cybersecurity firm Cyberhaven, one of whose employees was targeted by a phishing attack on December 24, allowing the threat actors to publish a malicious version of the extension. On December 27, Cyberhaven disclosed that a threat actor compromised its browser extension and injected malicious code to communicate with an external command-and-control (C&C) server located on the domain cyberhavenext[.]pro, download additional configuration files, and exfiltrate user data. The phishing email, which purported...

This week showed just how fast things can go wrong when no one’s watching. Some attacks were silent and sneaky. Others used tools we trust every day — like AI, VPNs, or app stores — to cause damage without setting off alarms. It’s not just about hacking anymore. Criminals are building systems to make money, spy, or spread malware like it’s a business. And in some cases, they’re using the same apps and services that businesses rely on — flipping the script without anyone noticing at first. The scary part? Some threats weren’t even bugs — just clever use of features we all take for granted. And by the time people figured it out, the damage was done. Let’s look at what really happened, why it matters, and what we should all be thinking about now. ⚡ Threat of the Week Silently Patched Fortinet Flaw Comes Under Attack — A vulnerability that was patched by Fortinet in FortiWeb Web Application Firewall (WAF) has been exploited in the wild since early October 2025 by threat actors to c...

A previously unknown vulnerability in OpenAI ChatGPT allowed sensitive conversation data to be exfiltrated without user knowledge or consent, according to new findings from Check Point. "A single malicious prompt could turn an otherwise ordinary conversation into a covert exfiltration channel, leaking user messages, uploaded files, and other sensitive content," the cybersecurity company said in a report published today. "A backdoored GPT could abuse the same weakness to obtain access to user data without the user's awareness or consent." Following responsible disclosure, OpenAI addressed the issue on February 20, 2026. There is no evidence that the issue was ever exploited in a malicious context. While ChatGPT is built with various guardrails to prevent unauthorized data sharing or generate direct outbound network requests , the newly discovered vulnerability bypasses these safeguards entirely by exploiting a side channel originating from the Linux runtime ...

Cybersecurity researchers have discovered more than half-a-dozen services advertisements for illegal access to artificial intelligence (AI) models on underground cybercrime forums and messaging platforms. One such service, Poison Claude, claims to offer access to Anthropic's large language models (LLMs), including Opus 4.8, Opus 4.7, Opus 4.6, and Sonnet 4.6. "Advertisements for Poison Claude explain how the service can offer the cheap tokens: by taking advantage of free bonus credits, such as the US$100 bonus credit on AWS for Bedrock accounts," Okta researchers Jeremy Kirk and Mathew Woodyard said in an analysis published Tuesday. "The service plainly states on its website that: 'We add those accounts to our pool, your request is routed to a specific account under the hood (you don't see this), and you get charged 5-15% of the official per-token price depending on the model.'" Poison Claude accepts payments in cryptocurrencies. Once a cus...

Rough Monday. Somebody poisoned a trusted download again, somebody else turned cloud servers into public housing, and a few crews are still getting into boxes with bugs that should’ve died years ago — the same old holes, same lazy access paths, same “how the hell is this still open” feeling. One report this week basically reads like a guy tripped over root access by accident and decided to stay there. The weird part is how normal this all sounds now. Fake updates. Quiet backdoors. Remote tools are used like skeleton keys. Forum rats swapping stolen access while defenders burn another weekend chasing logs and praying the weird traffic is just monitoring noise. The Internet’s held together with duct tape and bad sleep. Anyway, Monday recap time. Same fire. New smoke. ⚡ Threat of the Week Ivanti EPMM and Palo Alto Networks PAN-OS Flaws Under Attack —Ivanti warned customers that attackers have successfully weaponized CVE-2026-6973, an improper input validation defect in Endpoint Man...

This week, the shadows moved faster than the patches. While most teams were still triaging last month’s alerts, attackers had already turned control panels into kill switches, kernels into open doors, and open-source pipelines into silent delivery systems. The game has shifted from breach to occupation. They’re living inside SaaS sessions, pushing code with trusted commits, and scaling operations like legitimate businesses — except their product is chaos. And the underground is getting uncomfortably professional. Here’s the full weekly cybersecurity recap: ⚡ Threat of the Week cPanel Flaw Comes Under Attack —A critical flaw in cPanel and WebHost Manager (WHM) has come under active exploitation in the wild. The vulnerability, tracked as CVE-2026-41940, could result in an authentication bypass and allow remote attackers to gain elevated control of the control panel. In some cases , the attacks have led to a complete wipe of entire websites and backups. Other attacks have deployed ...

This week kept coming back to permission. A model crossed a boundary. A wallet trusted bad randomness. Webmail kept an intruder around. Public systems, package feeds, hotel networks, and login flows all gave away more than intended. Some of it was clever. Most of it was just access left lying around: old bugs, exposed gear, poisoned dependencies, weak defaults, and tooling that moved from forum chatter to real targets. The full weekly recap report follows. ⚡ Threat of the Week Anthropic Disclosed its Models Targeted 3 Organizations - Anthropic revealed that three of its models, including Claude Opus 4.7, Mythos 5, and an unnamed research model, breached three unnamed organizations during cybersecurity testing without its knowledge. The AI firm said the earliest incidents date back to April 2026, adding it made the discoveries after launching a "large-scale retrospective review" in response to the recent Hugging Face incident. "After reviewing 141,006 evaluatio...

Monday’s recap shows the same pattern in different places. A third-party tool becomes a way in, then leads to internal access. A trusted download path is briefly swapped to deliver malware. Browser extensions act normally while pulling data and running code. Even update channels are used to push payloads. It’s not breaking systems—it’s bending trust. There’s also a shift in how attacks run. Slower check-ins, multi-stage payloads, andmore code kept in memory. Attackers lean on real tools and normal workflows instead of custom builds. Some cases hint at supply-chain spread, where one weak link reaches further than expected. Go through the whole recap. The pattern across access, execution, and control only shows up when you see it all together. ⚡ Threat of the Week Vercel Discloses Data Breach —Web infrastructure provider Vercel has disclosed a security breach that allows bad actors to gain unauthorized access to "certain" internal Vercel systems. The incident originated f...

DALL-E for coders? That’s the promise behind vibe coding, a term describing the use of natural language to create software. While this ushers in a new era of AI-generated code, it introduces "silent killer" vulnerabilities: exploitable flaws that evade traditional security tools despite perfect test performance. A detailed analysis of secure vibe coding practices is available here . TL;DR: Secure Vibe Coding Vibe coding, using natural language to generate software with AI, is revolutionizing development in 2025. But while it accelerates prototyping and democratizes coding, it also introduces “silent killer” vulnerabilities: exploitable flaws that pass tests but evade traditional security tools. This article explores: Real-world examples of AI-generated code in production Shocking stats: 40% higher secret exposure in AI-assisted repos Why LLMs omit security unless explicitly prompted Secure prompting techniques and tool comparisons (GPT-4, Claude, Cursor, etc.) Reg...

OpenAI on Tuesday revealed the rogue artificial intelligence (AI) agent that escaped its sealed evaluation environment and broke into Hugging Face's production environment also hacked multiple third-party accounts and services as part of the attack. The latest disclosure shows that the security incident, which stemmed from an internal security test, was more extensive in scope than previously thought. The AI company said its ongoing review of the incident revealed a "small number of cases" where the models, including GPT-5.6 Sol and an "even more capable pre-release model," identified and used exposed credentials at the account-level on other publicly-available services. "This includes four accounts on four services as part of the Hugging Face incident (and a few accounts accessed as part of other evaluations)," it said. "One of these four accounts was used as an outbound relay and staging path, and another account was used for data st...