artificial intelligence | Breaking Cybersecurity News | The Hacker News

A malicious MCP server could trick an application built on the official  MCP Python SDK  into handing over the OAuth credentials it uses to log in to a real service, the SDK's maintainers said in a security advisory. Affected versions sent the client secret, the authorization code, and the PKCE proof key to a token endpoint the attacker controlled. The fix is in versions 1.30.0 and 2.2.0. The Model Context Protocol (MCP) is an open standard for connecting AI applications to outside tools and data, and this package is its official Python SDK for building MCP servers and clients. With the stolen credentials, the attacker can request a valid access token from the real login service. Cycode, the security firm that reported the flaw , demonstrated that full exchange in a test and says the resulting token carries whatever permissions the app was granted. The client secret is long-lived, so it keeps working until it is changed. The flaw is rated high (7.5) for the two prov...

OpenAI on Monday shelved plans to release GPT-6.1 Astra, a next-generation artificial intelligence (AI) model that was planned for an October launch, after it failed internal safety and alignment audits. The development was first reported by The Wall Street Journal. The move "marks a rare case of a major AI developer ditching a new release because of safety concerns," the news publication said. The ChatGPT maker said it made the decision to scrap its GPT-6.1 Astra model release after testing raised questions about whether it can follow user instructions without deviating from expected behavior. The Journal reported that the model exhibited higher levels of deception than its predecessor during evaluation, and failed to disclose what actions it had carried out. In some cases, it went ahead without seeking permission or attempted to use outside tools in scenarios where doing so could be deemed unsafe. "While (GPT-6.1 Astra) improved on axes such as model laziness...

OpenAI said it has made the decision to pause training of its most powerful models after one of its agents during reinforcement learning (RL) training contacted an external chatbot by exploiting a loophole in its internet-access restrictions. "An agent attempting to complete a search-based training task queried a public chatbot service through a gap in our internet-access restrictions: insufficient DNS filtering in its training sandbox," OpenAI said . "Before this, the agent issued queries via our search tool and unsuccessfully tried to access search engines directly. Note that all internet access apart from the DNS resolver in this report hit our offline webcache and therefore did not access the live internet." OpenAI said it has since added blocking controls at two independent layers to prevent this access in the first place. It also said its misalignment monitoring system detected the behavior within 15 minutes and it was acknowledged by a human reviewer...

What is IAM for AI agents? AI agents authenticate, invoke tools, and act across enterprise systems with delegated authority. IAM for AI Agents is the identity-control architecture that governs those actors. This guide covers the limits of conventional provisioning, the components that matter, how to evaluate framework choices, and what runtime evidence proves an agent behaved as intended. Identity and access management (IAM) for AI agents treats each agent as a non-human identity with a human owner, a defined purpose, scoped authorization, an expiration, and continuous monitoring. The complication is architectural. IAM platforms express intended access, while applications and infrastructure reveal what the agent actually executed. Between the two sits identity dark matter: the agents, credentials, application-local accounts, and authentication paths that central identity data never reports. A framework that cannot observe that surface produces policy intent, not assurance. Wh...

RatHat's operators build and publish the Android banking trojan and control infected phones from a web console, according to security company Cleafy. Cleafy has  traced nearly 100 deployments  of that console since April 2026. It said this fits a malware-as-a-service model, in which each customer runs a separate copy. The console stores what the malware collects from each phone, including text messages and passwords entered into fake login screens overlaid on banking apps. Its latest version asks Google's Gemini AI model to estimate each victim's bank balance from those messages and sorts the phones into high-value and mid-value groups. Nothing in the samples Cleafy analyzed uses the model to move money. Its role is "deciding which victims are worth an operator's time," the company said. One Console, Three Versions The malware on victims' phones has changed little since late 2025, Cleafy said. The console behind it has been replaced. Samples f...

A domain used as harmless placeholder text showed up in roughly 1,700 repositories. Then somebody registered it and started serving malicious lures. That is the kind of week this was: forgotten assumptions turning into live attack surface. Elsewhere, weak service accounts, old bugs, exposed systems, phishing kits, and strangely easy exploit paths kept doing useful work for attackers. Nothing exotic. Mostly things nobody expected to matter anymore. Here’s the full recap of what mattered this week. ⚡ Threat of the Week Citrix Warns of Actively Exploited NetScaler ADC and Gateway Flaws — Citrix released patches to address multiple vulnerabilities, including CVE-2026-88771 and CVE-2026-88772, that have come under active exploitation. CVE-2026-88771 is an improper input validation vulnerability that could allow an unauthenticated attacker to execute arbitrary commands, while successful exploitation of CVE-2026-88772 could allow for remote code execution or denial-of-service. CISA...

AI agents are moving into production faster than security teams can govern them. They are connecting to apps, handling data, calling APIs, and acting across business systems—often without the same controls applied to human users. According to Okta’s Global CISO Insights 2026 report, only 47% of CISOs are confident they can identify every AI agent in their environment. Even among those who feel confident about visibility, roughly 80% still worry that excessive access may be going unreviewed. That is the real problem: seeing an agent is not the same as controlling what it can do. Join Matt Immler , Regional CSO at Okta, for a practical session on how to bring AI agents under stronger identity governance before excessive access becomes harder to contain.

Cybersecurity researchers have disclosed details of a new botnet malware called Carbonato that's targeting exposed Docker daemons to deploy an open-source artificial intelligence (AI) agent framework called Hermes Agent . "The implant installs the framework unchanged, then overwrites its SOUL.md persona file," ThreatDown said . "The 39-line prompt directs it to execute tasks received through Telegram, maintain persistence, and collect credentials." At a high level, the botnet breaks into Docker daemons exposed without authentication on port 2375 and scans neighboring networks every five minutes to propagate further. On each host, it installs Hermes Agent with instructions to follow operators' Telegram commands. The cybersecurity company said it found the operation through an unauthenticated Docker registry that's been publicly accessible since May 2026. The staged data has been found to include details of the botnet and a separate campaign that d...

The way we talk about AI agents is shifting, and the way we implement them requires an even more fundamental shift. While earlier discourse focused on how quickly organizations could stand up agents and how much productivity they could promise, a string of recent incidents, including a widely discussed intrusion at Hugging Face during an evaluation of OpenAI agents, has spurred organizations to examine whether speed has outpaced the ability to secure what gets deployed . Security teams are increasingly asking what an agent can reach once it's running, and whether anyone would notice before it mattered. However tempting it may be to jump directly into enforcement controls and detections, you need to look before you leap. Research from Veeam reveals that 70% of organizations admit that AI workflows are already in contact with sensitive corporate data without full oversight in place, and 67% report that IT cannot fully track the autonomous workflows that employees are building. Sh...

Security leaders keep debating whether AI will produce an entirely new class of cyberattack. The nearer change is quieter and already visible: AI has made a failed attack cheap to retry. The routine version looks like this. An attacker lands on a low-privilege cloud account, and the first try at privilege escalation goes nowhere. That dead end used to cost hours of documentation reading, permission checks, and script debugging, and plenty of operators simply got stuck. With a model in the loop, the error gets explained, the script gets fixed, and a fresh enumeration path is under test within minutes. No step in that sequence is a new capability. Together they strip time, skill, and cost out of the unglamorous middle of an intrusion, the research and troubleshooting that sit between intent and outcome. What the threat reporting shows The public record traces the arc. In early 2025, Google's Threat Intelligence Group found state-backed actors treating generative AI as a prod...

This week, the dangerous stuff keeps arriving dressed as something boring. An update. A login box. A search answer. A coding tool. A link you have clicked a hundred times before. That is the thread running through the pile. Trusted paths get poisoned. Old bugs find new jobs. AI tools leak more than expected. Fake prompts look real enough. And some attacks barely need an exploit at all — just one weak setting or one person doing what the screen tells them. Nothing here looks especially dramatic. That is what makes it useful. The threats change every week. Subscribe, and we’ll alert you when each new ThreatsDay Bulletin is out.

The logistics sector has become the target of a new malicious cyber campaign that distributes an Android spyware codenamed Corp MDM . According to Have I Been Squatted, the campaign uses fake Google Play pages branded as CEVA and TKW Logistics to distribute an Android Package Kit (APK) file that's dressed up as a system service. The delivered app has the package name "com.corp.mdm" Corp MDM is a "compact surveillance implant designed to exfiltrate newly received SMS content, divert calls, and maintain a hidden foreground service," security researcher Ben Folland said . The malware has been described as narrow by design, lacking in spyware functions typically observed in commercial Android spyware. It's suspected that the threat actor behind the campaign used artificial intelligence (AI) during the development phase, given the presence of bugs that interfere with its capabilities. In addition, the activity is said to be part of a broader campaign t...

AI coding agents are changing how quickly developers can build and ship software as well as how quickly credentials can become exposed. According to GitGuardian’s 2026 State of Secrets Sprawl Report , commits identified as AI-assisted are leaking secrets at approximately twice the rate of human-written ones. Most of the fastest-growing categories of leaked credentials are now connected to AI services, meaning the tools meant to advance development are also accelerating the exposure of the keys development relies on. This isn’t a new vulnerability; what is new is AI changing the scale and pace at which those mistakes can happen. A coding agent can read an entire project, modify files, generate configurations and interact with external services in the time a developer might take to review a single pull request. The main issue isn’t that AI agents sometimes encounter secrets but that many of those secrets were never designed for an environment in which software can act autonomously. A...

An AI agent on an internal OpenAI research task bypassed access controls on an Australian government Medicare statistics portal in June, Prime Minister  Anthony Albanese said . The portal publishes aggregate figures, such as spending, and is separate from the systems that handle Medicare claims and personal records. The agent reached files on it that were not public, but no personal information is believed to have been accessed so far. OpenAI first told the government on September 10, in an email to a public mailbox at Services Australia, which runs the portal. OpenAI says it found the activity in August. Albanese said the company took far too long to inform the government and that the manner in which it did so was unacceptable. On June 18, the portal repeatedly refused the agent's data requests, but the agent found a workaround and gained unauthorized access. The government has not said how the agent got past them. Services Australia has told the government that the age...

A Windows malware called CLOSEDQUORUM is built to take orders from a vote of up to four AI models instead of an attacker's server, Cisco Talos said on September 22. The models can choose to steal Windows credentials, saved browser passwords, and crypto wallet data. Talos has not seen this setup work from start to finish, and the public version of the malware does not work as it is. Talos found the malware with  CAIRN , an open-source tool it released the same day to hunt for malware that uses AI services. The malware is at least three months old, because Talos's analysis of the code is dated June 17, 2026. The researchers did not describe how the malware would get onto a victim's computer. It said clues in the code tied the developer to criminal forum posts about carding, the trade in stolen card data, dating to 2025. How the AI Vote Works Malware usually takes orders from a command-and-control (C2) server that the attacker runs. CLOSEDQUORUM instead asks up t...

Autonomous security agents are getting good at finding bugs. Nobody has a good way to measure how good. Point one at a realistic target and what comes back is a report the agent wrote about itself: confident prose, a list of findings, and no way to tell which of them happened. Someone with a security background then sits down and checks every claim against the target. Which findings are real, which are duplicates, which are inventions, and, the question nobody has time for, what did the agent never try? That is a day of expert work for one run. Multiply it by three models, four prompt variants and ten repetitions, and the review queue is longer than the experiment. XRanges for AI, built by CTF.ae, exists for that loop. It deploys realistic target applications with instrumentation baked into every service, records what an agent actually does inside them, and scores each run live on four independent signals. This walkthrough covers how it works, what a run looks like from deployment t...

Anthropic and OpenAI on Tuesday announced new models, with both artificial intelligence (AI) companies noting that they are continuing to invest in improving alignment to combat risky behavior. Opus 5.5, per Anthropic , is a "major step up from Opus 5," and "achieves the best scores of any model to date on our automated behavioral audit, our alignment suite that tests Claude across thousands of simulated scenarios." The AI company said the model is less likely than its other recent models to carry out hard-to-reverse actions or act outside the boundaries it's been given, adding it's more resistant than Opus 5 to prompt injection. In its systems card, Anthropic explained that the latest model exhibited less misaligned behavior and less cooperation with misuse than any other recent Claude model. "It also took overeager or destructive actions less than any other model we tested," Anthropic said. "Regressions included being more likely ...

A use-after-free in the Linux kernel's AF_UNIX socket subsystem can be used to escape a container and gain root on the host, security firm DepthFirst  said in research published September 22. The flaw, tracked as  CVE-2026-80521  (CVSS score: 7.8), was fixed upstream on August 6, but Ubuntu has not shipped the patch for its 26.04, 24.04, or 22.04 LTS releases. DepthFirst  released exploit code  targeting Ubuntu 26.04. Ubuntu's security tracker  lists the Linux package on 26.04 as "vulnerable, work in progress." The 24.04 and 22.04 releases are also affected through newer kernel packages, including those for AWS, Azure, and GCP workloads. No fix has shipped on any affected release. The flaw is not in CISA's Known Exploited Vulnerabilities catalog, and there are no confirmed reports of attacks using it. The vulnerability sits in the kernel's garbage collector for AF_UNIX sockets. That collector cleans up file descriptors passed between processes t...

Microsoft on Tuesday announced the takedown of the EvilTokens device code phishing service that it said used artificial intelligence (AI) "at every step of the attack chain." The action, carried out with authorization from the U.S. District Court for the Eastern District of Virginia, involved the efforts of Health-ISAC, alongside Cloudflare, Coinbase, OpenAI, Railway, SpyCloud, The Shadowserver Foundation, and TRM Labs. Microsoft is tracking the threat actors behind the development and support of EvilTokens as Storm-2992 . In tandem, the Metropolitan Police Service arrested two men, aged 32 and 38, on September 11, 2026, in connection with the illicit commercial operation. The tech giant described EvilTokens as a "powerful cybercrime platform" that used AI to compromise email accounts and design roadmaps for financial fraud and scams. "While EvilTokens helped cybercriminals access email accounts, at the center of the service was an AI-style chatbot th...

A critical vulnerability in  Bifrost , an open-source AI gateway that routes requests to more than 20 LLM providers, allows an unauthenticated attacker to run arbitrary commands on the gateway server with a single HTTP request. The flaw, tracked as  CVE-2026-90898  (CVSS score: 9.8), affects all versions of the Bifrost HTTP transport before 2.1.0 when management authentication is disabled, which is the default configuration. A fix is available in transports/v2.1.0. Yuval Moravchick of  JFrog Security Research , who discovered the flaw, said an attacker can register a stdio-type MCP client through a single unauthenticated POST to the management API endpoint /api/mcp/client. Bifrost starts the specified command immediately, before any MCP handshake, as the gateway process user. On the official Docker image, that user is appuser. Because the gateway stores API keys for every connected provider, executing commands on the gateway process grants the attack...