AI attacks this week: A forgotten key let an attacker into Barracuda's cloud, and the traces match an AI agent (week 46/2026)

Alexander Busse·November 16, 2026

New articles by email

We send you an email whenever a new article is published. Only the new article, no advertising. You can unsubscribe at any time via the link in every email.

You will first receive a confirmation email. By subscribing you also consent to our email service Brevo recording whether you open the emails and click links in them. Details are in our privacy policy.

A robot stands in the open door of a small office building; a green key is still in the lock, and open door frames line up behind it.

Reporting period: 9 to 14 November 2026.

What happened this week

  • Barracuda: An old cloud key held by an AI vendor opened the way into Barracuda's own cloud. According to Barracuda, the traces match ARTEX.
  • ARTEX, sixth week in a row: The FBI warns in its own FLASH report. Unit 42 and Cloudflare describe intrusions from their own incident response work, and Chainstack now names ARTEX explicitly as the tool.
  • FaceStealer: The US Justice Department has seized infrastructure of a North Korean scheme that stole more than 400 million US dollars through fake video meetings.
  • Fireworks, Cursor, Notion: A key held by Fireworks led to agent logs containing credentials and to a tampered extension for up to 5,000 developers. Railway puts the misuse of stolen AI keys at about 740,000 US dollars.
  • Mittwald: The German hosting provider was attacked through its MCP server. On 12 November the attackers struck again.
  • Cairn: According to Unit 42, the AI framework ran 25 operations against at least 19 organisations on its own.
  • AI as a tutor: Symantec describes a stalled intrusion that got going again with AI-generated commands. In Spain, the Guardia Civil has arrested a man who allegedly relied heavily on AI tools for his intrusions.
  • Vulnerabilities: The US agency CISA lists a flaw in GPUStack as exploited. NEAR AI was attacked through a flaw in SGLang.

Barracuda: A forgotten key and 44 hours

Barracuda, a vendor of security products, published a customer advisory on 10 November and a detailed analysis by its chief information security officer on 11 November. The attacker came in with a valid cloud key. Barracuda had given it to a vendor for an earlier project, "a company working in AI". When the work ended, Barracuda did not retire the key: "That is on us." In October the vendor's development environment was compromised. According to Barracuda, the key sat there in a configuration file in the vendor's GitHub repository.

On 1 November at 04:33 UTC the attacker used the key for the first time. Within minutes it was mapping what the key could reach and reading stored secrets. Each secret it found led to the next system. The path ran from the cloud accounts to a data analytics system, then to a source code token and on to Kubernetes. In the early hours of 2 November it spent a little under four hours creating short-lived branches in a handful of repositories. This made a build workflow write the build system's secrets into the log. The secrets were encoded so that the usual masking in the log would not hide them. Barracuda recorded the last activity on 2 November at 23:31 UTC. In all, the attacker reached about ten cloud accounts in roughly 44 hours. It searched specifically for crypto wallets and left a container with read-only access to a server's file system as a backdoor.

On 4 November an engineer noticed unusual activity on their own account. On the same day the vendor got in touch about its own incident. "By the time the activity was reported, it had already stopped." Barracuda found no direct evidence that the attacker viewed customers' credentials for Barracuda XDR. According to Barracuda, it cannot rule this out because the logging was not detailed enough. Barracuda notified the potentially affected customers directly and asked them to revoke and recreate their integration credentials.

Barracuda draws this lesson from the incident: "Detect-and-respond, as the primary strategy, is dead." According to Barracuda, short-lived, narrowly scoped credentials would have stopped the intrusion at its first step. Credentials held by vendors are to expire when their engagements end.

Role of AI: According to Barracuda, the names of the resources created, the order of the steps and the built-in routines for hunting credentials and wallets match ARTEX. In ARTEX a planning AI agent decides the next step, and other agents carry it out. Because many groups use ARTEX, Barracuda says the match allows no conclusion about who the attacker was.

How well documented: Barracuda confirmed the incident itself and published a timeline with times of day. The link to ARTEX rests on features of the traces; Barracuda does not report finding the tool itself. Barracuda does not name the vendor either.

ARTEX in its sixth week

The FBI published a FLASH report on ARTEX on 13 November. It says a variety of actors have used ARTEX since at least August 2026. Some intrusions went from initial targeting to privileged access in under three hours. Besides the open-source version there is a commercial one called ARTEX PRO. According to the FBI, the attackers most often use language models from the DeepSeek Flash family. In its footnotes, the report supports its findings with publications by vendors such as Microsoft, CrowdStrike and Zscaler.

Unit 42 describes intrusions from its own incident response work. Attackers using ARTEX turned public password reset pages and exposed APIs into a path to cloud access, sometimes in less than a day. According to Unit 42, they did not need to exploit a single software vulnerability. Cloudflare counts more than 50 families of such frameworks that tried to scan Cloudflare customers worldwide in the three days before its report of 13 November. Since the original ARTEX repository was removed on 8 October, the traffic from these tools measured by Cloudflare has grown more than tenfold. In its own incident response work, Cloudflare repeatedly saw attackers take over static, non-expiring tokens in CI/CD pipelines.

Follow-up to week 45: Chainstack added technical details to its incident page on 10 November and now writes: "The attacker used ARTEX". The way in was an internet-exposed Grafana instance in a development Kubernetes cluster whose default administrator credentials had never been changed. From there the attacker moved through an MCP server without authentication whose service account had administrator rights for the whole cluster. Chainstack found no evidence that a software vulnerability was exploited. In week 45 we wrote that Chainstack did not show that AI had been used. The update makes that statement out of date.

Role of AI: In ARTEX, AI agents plan the steps of an intrusion and carry them out. According to Unit 42, humans were still not entirely absent from the attacks it observed. In one case the attacker switched from scripts to a browser to reset the password.

How well documented: The FBI summarises reports by vendors and by CERT-EU. Unit 42 and Cloudflare rely on their own incident response work, Chainstack on its own investigation. As one example, the FBI cites attacks on South Korean financial institutions in which personal information of about 67,000 people was exfiltrated. It bases this on reports by CrowdStrike, Zscaler and CERT-EU. Unit 42 and Cloudflare name neither victims nor amounts of damage.

FaceStealer: Fake video meetings bring North Korea more than 400 million US dollars

The US Department of Justice announced actions against North Korea on 12 November. Among other things, it seized cloud infrastructure of the FaceStealer scheme, which has been used to steal cryptocurrency worth more than 400 million US dollars. The attackers send invitations to video meetings from accounts stolen from earlier victims. In the meeting they record the target's webcam video and play recorded videos of earlier victims the target knows. This makes it look as if trusted contacts are taking part. The FBI seized more than 900 webcam videos of victims. Recently the scheme has increasingly targeted people who develop and maintain open-source software. According to the department, this appears calculated to enable attacks on the software supply chain.

In the same move, the department seized or restrained more than 1.2 billion US dollars in cryptocurrency from the theft at Binance. We covered the Binance case in week 45.

Role of AI: According to the department, the perpetrators use "super intelligence (SI)" to create deepfake videos of well-known and influential people. Arctic Wolf writes that generative AI supplies the attackers with artwork, avatars and realistic synthetic video of meeting participants. In Arctic Wolf's view, the core of how the intrusion works has stayed the same.

How well documented: The details come from the US Department of Justice. Arctic Wolf places the scheme in context in its own report. The groups behind it are publicly known under names such as BlueNoroff and Sapphire Sleet.

Fireworks, Cursor, Notion: How one key set off a chain

On 13 November Cursor described how the incident at the end of October unfolded. "It started at Fireworks." An attacker obtained a key that Cursor had issued to the AI provider Fireworks. With it the attacker copied data from a storage bucket Cursor used to share data with Fireworks. It held coding agent traces, mostly from 2025, from sessions not in Privacy Mode. Some of them contained credentials that had not been redacted. Using a stolen credential from the exposed data, the attacker placed tampered copies of Anthropic's Claude Code extension on the cache that serves Cursor's extension marketplace. At most 5,000 users may have received the tampered build on 29 October. According to Cursor, Anthropic did not publish this copy, and there was no vulnerability in Anthropic's product.

In an update dated 9 November, Fireworks writes that incomplete model weights and datasets of some customers were accessed. It describes the attackers as financially motivated and aiming to sell unauthorised access to language models, known as LLMjacking. Notion adds that the exposed data included Notion AI chat transcripts used for the Insights feature. On 11 November Notion revoked exposed Notion API keys.

On 10 November Railway published its final report on the October break-in, which we covered in week 43. According to Railway, keys for AI providers were the main target. The stolen keys caused an estimated 740,000 US dollars in unauthorised spend. Railway describes the defence as a contest of agent speed. According to Wiz, the attackers used models with their safety refusals removed and a tool similar to ARTEX.

Role of AI: Here AI is mainly the target. Fireworks names the resale of access to language models as the motive, Railway names keys for AI providers as the main target. At Cursor, the coding agent traces supplied the credentials for the next step. Only Railway, citing Wiz, says the attackers themselves worked with AI agents.

How well documented: Cursor, Fireworks, Notion and Railway confirmed their incidents themselves. None of the companies names a perpetrator.

Mittwald: In through its own MCP server

On 7 November the German hosting provider Mittwald, based in Espelkamp, described in its blog how an attacker got into its platform. Mittwald runs an MCP server through which customers could connect their mStudio management interface to AI assistants. According to its FAQ, access tokens of users of this interface were read through it. On 27 October at 14:58 the attacker registered another account and at 15:03 deliberately exploited the flaw in the MCP server. Mittwald forensically analysed 56 compromised projects and the MCP server and informed the relevant authorities.

On 12 November Mittwald reported on its status page that the attackers had resumed their attacks. They exploited a security flaw and compromised the company website, which Mittwald took offline. By then Mittwald had found no access to internal systems. Because Mittwald considers the attackers to be professional actors using automated AI systems, the provider switched off its customer centre, mStudio and all SSH connections as a precaution. Websites hosted at Mittwald remained reachable. On 13 November the interfaces came back online. According to Mittwald, there is no longer any outside access to its systems.

Role of AI: The MCP server, an interface for AI assistants, was the way in. Mittwald infers from the speed that the attacker used AI: at times only seconds passed between steps that would take a human much longer.

How well documented: Mittwald confirmed the incident itself and informed the authorities. The use of AI tools is Mittwald's assessment; beyond the speed, the provider gives no evidence for it. Mittwald has announced a full post-mortem.

Cairn: Target and goal from humans, the rest done by AI

Unit 42 describes a group it calls CL-STA-1238. The group had the open-source framework Cairn run 25 operations autonomously against at least 19 organisations, including in the semiconductor, defence and civil society sectors. "The attackers supplied a target and a goal, then Cairn planned, executed and recorded results on its own." The operations lasted from 20 minutes to more than 50 hours. At ten organisations the attackers achieved access or retrieved data, at one of them more than one million records. The instructions to the model declared every task authorised, and each operation created a file declaring the engagement authorised. Neither Zhipu GLM nor DeepSeek refused this. Unit 42 assesses with moderate confidence that the activity is China-nexus intelligence collection.

Role of AI: Cairn planned and carried out the operations, humans supplied the target and the goal. Unit 42 sums up: "Autonomous AI raises the volume and coverage of exploitation, but not its complexity."

How well documented: The only source is the analysis by Unit 42. Unit 42 does not name the organisations, and the attribution is an assessment with moderate confidence.

AI as a tutor

Symantec describes an intrusion at a US retailer in September. At first the attackers typed their commands by hand: 110 actions with three typing errors, each corrected within seconds. They re-ran their blocked loader five times in about 20 minutes without changing their approach. On 23 September around 50 long scripts followed in roughly two hours, none with a typing error. Symantec assesses that they came from an AI. Within minutes the attackers were harvesting credentials and moving laterally. Symantec has investigated dozens of AI-assisted attacks since May 2026. None of the techniques in this second phase were new.

In Spain, the Guardia Civil announced on 12 November the arrest of a man in Ciudad Real who allegedly made intensive use of AI tools for his intrusions. He used them to acquire specialist knowledge, solve technical problems, work out his procedures and get past obstacles, the Guardia Civil says. According to the Guardia Civil, his technical skills were far above what his training would suggest. Among those affected was the city of Ciudad Real, which had to suspend the platform for its low emission zone after the attack.

Role of AI: In both cases AI helped the attackers past obstacles. At the US retailer it supplied scripts without typing errors after the attackers had stalled. In Ciudad Real, according to the Guardia Civil, it supplied specialist knowledge and solutions to technical problems.

How well documented: Symantec relies on its own investigation. Symantec infers AI assistance from the form and error-free state of the commands with high confidence, and a human at the controls with moderate confidence. The details of the arrest come from the Guardia Civil.

Vulnerabilities and entry points in AI infrastructure

  • GPUStack (CVE-2026-102895): The US agency CISA added the flaw to its catalogue of known exploited vulnerabilities on 9 November. GPUStack is a platform for running AI models, and the flaw allows administrator access without authentication. Severity critical (CVSS 9.8).
  • NEAR AI (report): Attackers exploited a remote code execution flaw in SGLang, the inference software running on NEAR AI's servers. They downloaded account details of about 100 users and reached 18 agents on a legacy deployment outside the protected computing environment. NEAR AI notes about the protected environment: "It does not protect against a vulnerability in the software running inside the environment." NEAR AI gives no identifier for the flaw.
  • Microsoft, Patch Tuesday on 10 November: Microsoft says it has fully mitigated a critical flaw in Azure OpenAI (CVE-2026-119990, CVSS 9.8), and customers need not act. In Microsoft MCP Gateway, an authenticated attacker could escalate their own privileges (CVE-2026-109492, CVSS 8.8). Here, according to Microsoft, customers must act themselves.

How well documented: CISA confirms the exploitation of GPUStack; who is exploiting the flaw is not known. NEAR AI confirmed its incident itself. Microsoft published its flaws itself and reports exploitation for neither.

What follows in our view

  • End keys together with the engagement. At Barracuda and at Cursor, the intrusion began with a key held by a vendor. Access for vendors should be time-limited and revoked when the project ends.
  • Short-lived credentials instead of permanent keys. Cloudflare repeatedly saw attackers take over non-expiring tokens in CI/CD pipelines. Barracuda considers short-lived, narrowly scoped credentials the step that would have stopped its own intrusion.
  • Close default passwords and exposed test environments. Chainstack was attacked through a Grafana instance with unchanged default credentials, and ARTEX needed no software flaw there.
  • Run MCP servers only with authentication and narrow rights. At Mittwald its own MCP server was the way in, at Chainstack an MCP server without authentication and with administrator rights.
  • Confirm invitations to video meetings through a second channel. With FaceStealer, the invitations came from hijacked accounts of known contacts, and recorded videos played in the meeting.

Selection and sources

We include reports in which AI was essential to the attack, as a tool, as a target or for deception. Every statement is checked against the linked source, usually the primary source. Where only media reports are available, we say so. We distinguish whether an incident has been confirmed by the affected party, an authority or the vendor, or has so far only been reported. We welcome notes on errors and make corrections visible.

Get new articles by email. Only the new article, no advertising.

New articles by email

We send you an email whenever a new article is published. Only the new article, no advertising. You can unsubscribe at any time via the link in every email.

You will first receive a confirmation email. By subscribing you also consent to our email service Brevo recording whether you open the emails and click links in them. Details are in our privacy policy.

Related articles

Related services