Get a demo — 30 minutes →
← Back to blog
Anomity robot illustrating This Week in AI Agent Security: Stealers, Agentic Breaches, and Sandbox Escapes
Insights

This Week in AI Agent Security: Stealers, Agentic Breaches, and Sandbox Escapes

TL;DR
  • VOLTA, a Windows infostealer disclosed by Cisco Talos, feeds stolen credentials into Claude Opus 4.6 (with Kimi K2 and Gemini 2.5 Flash fallbacks) to rank victims by monetization value, then exfiltrates ranked dossiers over Telegram.
  • Chainalysis reported an ongoing agentic cyber attack that exposed names, usernames, password hashes, and live tokens for all users, forcing KYT, Reactor, and related services offline. Microsoft and others tie related activity to frameworks such as ARTEX.
  • Hunt.io and CrowdStrike research on ARTEX shows a China-developed AI pentest framework used against South Korean financial firms, with the operator running Claude Code plus custom skills and routing inference through DeepSeek, GLM, and Grok.
  • Check Point detailed a now-patched ChatGPT Lockdown Mode bypass via HTML preview popups. Separately, Forever Security's ThisBash shows a malicious browser extension can drive Claude Desktop through Dispatch, including across devices on the same account.
  • OpenAI launched GPT-6 Cyber (Critical under its Preparedness Framework) for approved Daybreak Red customers: 81 novel high-severity Firefox findings and 948 Linux privilege-escalation exploits in disclosed research, with stronger in-scope alignment than GPT-5.6 Cyber.
  • SSHole shows OpenSSH on musl (Alpine and related images) accepting any password of 257 to 1024 characters for *-locked accounts when PAM is off. The finder used agent swarms. Everyday campaigns also impersonate Claude and ChatGPT installers to steal sessions.
  • The durable control for security teams is inventory of agents, MCPs, skills, extensions, and secrets on the endpoint, plus allow/deny/log at the agent hook before a call runs.

The first week of November 2026 compressed several years of agentic risk into a few disclosures. An infostealer now ranks victims with frontier models. A blockchain analytics firm shut key products after an agentic breach. An AI pentest framework left its own working files open. A browser extension can drive a desktop coding agent across devices. And the same week, OpenAI shipped a Critical-rated cyber model while researchers published a decade-old OpenSSH edge case found with agent swarms. None of these stories is only about a model prompt. Each one is about the agents, skills, extensions, MCPs, and secrets that already sit on endpoints - the layer traditional controls were never built to inventory.

This roundup sticks to verified public reporting. For the wider map of how these pieces fit together, see agentic AI attack surface by layer, GhostJacking, and LLMjacking.

VOLTA: an infostealer that ranks victims with Claude, Gemini, and Kimi

Cisco Talos, as covered by The Hacker News on 2 November 2026, described VOLTA-AUTODOXER (VOLTA), a Windows infostealer that harvests Chromium passwords and cookies, Outlook and Slack data, payment cards, cryptocurrency wallet seed phrases, Discord, Telegram, Steam, and Roblox sessions, Microsoft OAuth tokens, SSH and cloud-provider keys, Wi-Fi profiles, and screenshots. Talos researcher Ryan Fetterman said VOLTA builds a structured evidence bundle about the person behind the machine - geolocation and home address, employer, inferred birth year, card last-four digits, Slack workspaces, document authorship metadata, iPhone backup indicators, and a live Roblox account valuation - then uses stolen session access to search Gmail for receipts from Amazon, DoorDash, Uber Eats, Grubhub, Walmart, and PayPal.

That bundle is serialized and sent through OpenRouter with a roughly 240-word system prompt that casts the model as a forensic intelligence analyst. Requests default to Anthropic's Claude Opus 4.6, falling back to Moonshot AI's Kimi K2 and Google's Gemini 2.5 Flash when a response contains refusal phrases such as "I can't" or "I'm not able." The required output includes a "WHAT THEY'RE WORTH" ranking and an "OPSEC" assessment. Exfiltration happens over Telegram in three stages: raw steal, AI dossier, then Gmail enrichment. Build-time templates with six slots (Telegram channel, API keys, settings) suggest a builder kit or bespoke compiles per customer. Talos's framing is blunt: the AI layer is effort displacement, not speed magic - any customer can point the tool and receive ranked assessments.

Talos also detailed NIK C2, a Go-based Windows remote access tool with more than 52 commands and an OpenAI client that will use a planted API key or, if none is supplied, scan the victim environment for OPENAI_API_KEY. That is textbook LLMjacking: the victim funds the attacker's inference, and the billing trail points at the compromised account. Zscaler ThreatLabz, quoted in the same coverage, notes GenAI is well suited to classify and search large volumes of stolen data once a host is compromised - the same triage pattern VOLTA productizes.

Chainalysis, ARTEX, and agentic offensive tooling in the wild

On 6 November 2026 The Hacker News reported that Chainalysis said its ongoing investigation into an agentic cyber attack also exposed names, usernames, password hashes, and live tokens for all users. While investigating, the company removed access to Reactor and Data Solutions, continued to report no evidence that investigations data was accessed or that attribution data was manipulated, and took Know Your Transaction (KYT) UI and API offline along with other address screening products and the login page. An earlier on-path attack had exposed a subset of customer passwords; that vector was addressed and impacted accounts locked. Restoration of KYT and Address Screening was planned behind trusted-IP allowlisting, with mandatory password resets and API key rotation.

Chainalysis said the activity resembled attacks on other tech platforms in recent weeks, naming Chainstack, OpenCode, Warp, and WorkOS, and referencing autonomous penetration testing tooling such as ARTEX. Microsoft, in the same reporting thread, described threat actors using an agentic AI offensive framework to search public GitHub repositories for exposed secrets and to conduct follow-on attacks, cutting hands-on-keyboard time. In at least one case, an AI agent reportedly progressed from scanning GitHub for secrets to exfiltrating data from Azure Key Vault in about five minutes, with activity assessed as ongoing from August 2026. Sign-in events from 199.19.77.84 and 199.19.77.18 were associated with service-principal abuse across tenants, including storage keys, App Service settings, AI and ML connections, and AKS credentials, with subsequent SharePoint, OneDrive, and mailbox access.

Google and Wiz reporting in the same period described an emerging LLM mixer market: illicit intermediaries that advertise access to compromised enterprise LLM accounts. Two resale operations named in coverage, llm-routing.com (momo) and xingjiabiapi.org, offered frontier models at discount prices using stolen AWS Bedrock keys. The IP 23.106.46.239, linked to llm-routing.com, hosted an Artex landing page in mid-October 2026 and was flagged after the Zeabur breach that exfiltrated over 3.2 million environment-variable records and data for 321,810 accounts. Stolen AI API keys are no longer only a cost problem; they are inventory for a resale market. That is why secrets management for AI agents and endpoint discovery of plaintext keys matter before the mixer lists your account.

ARTEX operators leave Claude Code skills and model routers in the open

A 5 November 2026 Cyber Security News report summarizing CrowdStrike, Genians, and Hunt.io research gives a concrete picture of ARTEX in a financially motivated campaign. Between late September and early October 2026, seven South Korean financial institutions had data stolen through internet-facing systems. Roughly 66,000 to 67,000 people and about 2,200 companies were affected, with names, jobs, income details, and loan information exposed. CrowdStrike judged with moderate confidence that a single Chinese-speaking, financially motivated lone operator used ARTEX rather than a known named group. The ARTEX GitHub page and Docker image were taken down on 11 October 2026, though copies remain online.

Hunt.io, pivoting from a known console host, found an open directory on a Hong Kong jump server whose state file showed the operator ran Claude Code version 2.1.221, started 510 times since 4 August 2026. Requests were not going to Anthropic: a Claude Code Router and helper scripts sent traffic through the xcai.pro reseller to DeepSeek, with additional Zhipu AI keys for GLM models and a grok2api tunnel - matching CrowdStrike's observation that DeepSeek, GLM, and Grok were in use. The state file listed 19 custom skills used 147 times, covering reconnaissance, code review, web testing, firewall bypass checks, Android analysis, and cloud phone control. One skill tested payment logic next to files holding test accounts for a cash-in deposit service, consistent with IDOR hunting. A .secrets file labeled in Chinese as the sole credential source fed reverse SSH tunnels and memory-file sync scripts.

Hunt.io also described hunting pivots: certificate search returning 19 IPs, malware-tag and console-title searches returning on the order of 1,300 hits each, and 280 open-directory results, with tags spanning about 30 countries led by the United States, China, and Hong Kong. The defensive lesson is not only "block ARTEX." It is that coding agents plus custom skills plus model routers are now the operator's workbench - the same artifact types agent skills governance and least privilege for agent skills exist to constrain on your own fleet.

ChatGPT Lockdown Mode bypassed through the app around the model

Check Point, via The Hacker News on 5 November 2026, disclosed a now-patched flaw that bypassed Lockdown Mode in OpenAI ChatGPT. Lockdown Mode shipped in June 2026 to limit outbound network requests from HTML page previews inside code blocks. Check Point found that nested sandboxed frames still allowed popups to escape the sandbox: by creating another nested frame, an attacker could reach an unmodified popup function and open content no longer subject to Lockdown Mode restrictions. Impact included silent chat data theft, phishing, and malware delivery, including zero-click exfiltration in the ChatGPT desktop app. OpenAI addressed the issue on 18 September 2026. Eli Smadja of Check Point's framing is the one worth keeping: AI security controls are only as strong as the wider application environment around them.

The same component class showed up again. Parameter Labs described a separate path where a shared ChatGPT conversation with a previewable HTML block could execute as chatgpt.com with the victim's authenticated web session after one click. Check Point said the two shortcomings are different bugs on the same surface. Adjacent research named in the same coverage includes SkillJacking, Plugin4Shell, BragJack, Embrace (embedded-browser issues across Cursor, Copilot, Codex, VS Code, and Claude Desktop), Repeat-After-Me against several agent frameworks, AgentCorruption on AWS Bedrock AgentCore, Cryptographic Context Injection against Grok and Copilot CLI, SalesBleed against Salesforce Agentforce, PleaseFix against agentic browsers, and Workflow Identity Hijacking. The pattern is consistent: the harness and the UI, not only the model weights, decide what untrusted input can reach.

ThisBash: a browser extension drives Claude Desktop across devices

Forever Security researcher Gal Weizman published ThisBash on 4 November 2026 (also covered by Cyber Security News and The Register). Claude Dispatch lets a user send tasks from phone or browser to Claude Desktop. Dispatch's interface and API were reachable from an ordinary logged-in claude.ai session; the web UI hid the link without blocking the route. An extension with scripting permission on claude.ai could open Dispatch in a tiny background window, fetch the desktop agent session ID, and send prompts as the user. Four handwritten prompts plus a one-second loop clicking Allow got Claude to install an npm package whose preinstall hook opened Calculator. When Claude hesitated, prompt forcing - sustained control of the chat channel, not a one-shot injection - supplied reassurance in the user's voice until the model complied.

Because Dispatch is cross-device by design, the blast radius scales: same machine equals browser-to-OS escape; another machine on the same account equals remote code execution; a personal Claude account signed into a work laptop equals an organization entry path. Weizman reported the issue to Anthropic on 14 May 2026; Anthropic confirmed validity on 19 May and closed it as a duplicate of an internal finding, with no CVE and no bounty reported. On 13 October 2026 Anthropic said Dispatch would be retired targeting end of October; as of 30 October the support page still described limited beta for existing Pro, Max, and Ultra users. Weizman argues the deeper design issue - treating an authenticated web session as authorization for local agent actions - also appears in Claude Code Remote Control and Cowork sessions. That is the same trust boundary our Claude Cowork integration post is written to close from the sensor side: visibility inside the place the agent actually runs.

GPT-6 Cyber raises the defensive and offensive ceiling

OpenAI announced GPT-6 Cyber on 5 November 2026 (Cyber Security News summarized the same day; numbers below match OpenAI's own post and system card). It is a cyber-specialized model for authorized vulnerability discovery, exploit development, and related work, initially available through the API and Codex to a select group of approved Daybreak Red customers, with expansion planned. OpenAI classifies it as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber.

Reported evaluation and research results from OpenAI: 48.5% on ExploitGym versus 32.1% for GPT-5.6 Cyber; 40.3% recall on a new web testing task versus 32.0% for GPT-6 Astra; 95% task completion on RemoteExploitBench while Astra with Daybreak Blue refused all tasks. Alignment: 99.93% in-scope on ExploitGym Honeypot without deployment safety prompts (up from 94.49% for GPT-5.6 Cyber), and 100% with those prompts. Real-world research claims include 81 previously unknown high-severity Firefox vulnerabilities (Mozilla advisory MFSA 2026-118 covers related Firefox 159 fixes) and 948 working Linux local privilege-escalation exploits from a pipeline that reproduced 1,724 of 2,308 candidate findings, with 304 critical findings disclosed with patches. Daybreak Red API access requires IP allowlisting by default; OpenAI also describes cyber-specific misuse monitoring, Private Safety Processing for zero-retention customers, and a Daybreak Red for Critical Infrastructure track plus a $1 billion Daybreak for Frontline Defenders grant program. Dragos, in a mock power-plant evaluation quoted by OpenAI, reported the model separating OT assets from IT noise and finding two zero-days in engineering software that handles protective relay data.

For security teams the governance question is unchanged from Daybreak defensive agent governance and autonomous hackbots: a Critical-rated cyber agent is still an agent with repository reach, secrets access, and code execution. Who may run it, on which machines, with which keys, and under which allow/deny policy is an endpoint and identity problem - not only a provider access-tier problem.

SSHole: agent swarms meet a ten-year musl edge case

The site sshole.com, published by Ozgun Kultekin of Wakefield, documents a behavior in OpenSSH on musl Linux (Alpine, OpenWrt, Buildroot, Yocto, and many Docker images) when password authentication is enabled and PAM is not used. Accounts whose password field is only * - including key-only deploy users, LDAP users, and root on Alpine after the 2022 switch from ! - accept any password between 257 and 1024 characters if another account on the host holds a real password hash. musl's crypt returns * for passwords over its 256-byte SHA-512 crypt cap; OpenSSH's timing-equalization path for invalid users can reuse another account's salt, so crypt(long_password, that_salt) returns *, which matches the locked account's stored field. Alpine's official AWS and Azure images ship a default alpine user with passwordless doas, so the path is effectively root. glibc, PAM, Dropbear, and OpenBSD are described as unaffected. OpenSSH 10.6p2 fixes it; disabling PasswordAuthentication is the interim control. Kultekin reports finding the bug with agent swarms throwing thousands of weird logins at builds across glibc, musl, PAM, and shadow configurations - another data point that autonomous research agents already change who finds what, and how fast.

Everyday enterprise campaigns still ride AI trust

ANY.RUN's October 2026 roundup, published via The Hacker News on 5 November, lists five active campaigns against US enterprises that security teams should not ignore while reading agent research:

  • SkyW4lker - Microsoft 365 phishing via legitimate device-code authentication; 618 related sandbox analyses between June and 21 October 2026, 46% from the US.
  • Fake AI campaigns - fraudulent ChatGPT and Claude apps and event invitations; 1,258 related sandbox analyses, with the US ranking second worldwide. In one case a genuine Claude installer ran while hidden processes stole browser session cookies.
  • GeorgeGinx - Phishing-as-a-Service redirecting Adobe Shared Links to Microsoft device-code pages against government, banking, healthcare, and IT targets.
  • WKLP - a Python infostealer and RAT observed against government organizations in the US and France that adds Defender exclusions, terminates security processes, and creates disguised firewall rules.
  • Mir0Auth - Microsoft 365 token theft via device-code flows, often starting from phishing PDFs that redirect to fake DocuSign pages.

Fake Claude and ChatGPT installers are the same class of problem as SEO-poisoned Gemini CLI and Claude Code installers: the install path itself is the attack surface, and the loot is the sessions and tokens the AI tooling already holds.

What security teams should do this week

  • Patch and harden SSH on musl. Upgrade OpenSSH to 10.6p2 on Alpine and related images, or set PasswordAuthentication no until you can. Audit * and ! locked accounts and cloud image defaults.
  • Inventory the agent layer. List agents, CLIs, MCP servers, skills, plugins, extensions, hooks, and secrets on every managed endpoint. Flag Claude Code installs, custom skills, personal AI accounts with cross-device remote control, and plaintext API keys.
  • Treat browser extensions as endpoint software. Restrict install rights, review high-permission extensions, and assume anything that can script claude.ai or chatgpt.com can influence connected desktop agents.
  • Rotate AI and cloud credentials. Assume keys that lived in env files, .secrets, or breached third parties are burned. Prefer short-lived tokens; watch for LLM cost spikes and new model traffic (LLMjacking).
  • Tighten M365 device-code auth. Restrict unused device-code flows in Entra Conditional Access; revoke suspicious sessions; treat unexpected DocuSign or "AI event" mail as hostile.
  • Govern cyber-capable models as high-privilege agents. Daybreak Red and similar access need SSO, MFA, IP allowlisting, sandboxes, and human review - the same posture you would give a privileged engineering workstation.
  • Assume the UI around the model is in scope. Lockdown Mode-style controls fail when popups, shared chats, or extensions bypass them. Test the app, not only the model card.

How Anomity helps govern the agent layer

Every item above shares a precondition: you have to see the agents, skills, extensions, MCPs, and secrets on the endpoint before you can govern them. Anomity runs a lightweight, unprivileged Endpoint Sensor that inventories eight AI artifact types - agents, MCP servers, extensions, plugins, skills, secrets, hooks, and CLIs - on Windows, macOS, and Linux. A Browser Sensor covers the AI people use in a tab and the accounts they sign in with. Cloud discovery surfaces OAuth grants against Google Workspace and GitHub. Secrets are redacted on the endpoint; only metadata leaves over HTTPS.

On agents that expose a hook, such as Claude Code PreToolUse, Anomity returns allow, deny, or log before a tool call runs (runtime governance). That is the control that matters when a custom skill, a stealer-driven triage loop, or an extension-driven Dispatch session tries to install a package, read a token store, or reach an unfamiliar outbound host. Violations route to SIEM, Slack, email, and Jira. Every added, removed, or modified artifact lands in a queryable 90-day audit trail. Anomity is SOC 2 Type II and complements Network, EDR, DLP, and GRC - it covers the agent layer those tools were never built to see.

You can't govern what you can't see.The Anomity principle

If you want that inventory and hook-level governance across your fleet this week, book a 30-minute demo.

Frequently asked questions

What is VOLTA and why does the AI layer matter?

VOLTA (also called VOLTA-AUTODOXER) is a Windows infostealer described by Cisco Talos that harvests browser passwords and cookies, Outlook and Slack data, payment cards, cryptocurrency seed phrases, messaging and gaming sessions, Microsoft OAuth tokens, SSH and cloud keys, Wi-Fi profiles, and screenshots. After building an evidence bundle, it sends that bundle through OpenRouter to Anthropic's Claude Opus 4.6 by default, falling back to Moonshot AI's Kimi K2 and Google's Gemini 2.5 Flash when a model refuses. The model returns a ranked monetization dossier. The AI layer does not invent a new theft path; it displaces the operator's triage work so any customer of a builder kit can receive prioritized victims in Telegram.

What did Chainalysis disclose about the agentic attack?

On 6 November 2026 reporting, Chainalysis said its investigation into an ongoing multi-front cyber attack that used offensive agentic techniques also exposed names, usernames, password hashes, and live tokens for all users. It removed access to Reactor and Data Solutions, took Know Your Transaction (KYT) UI and API offline along with address screening products and the login page, and said it had no evidence that investigations data was accessed or that attribution data was manipulated. Customers were urged to reset reused passwords elsewhere, then later to supply trusted IPs for allowlisting before rotating passwords and API keys.

How is ARTEX being used in real attacks?

CrowdStrike tied late-September to early-October 2026 breaches of seven South Korean financial institutions to ARTEX, an open-source China-developed AI penetration testing framework, with moderate confidence that a single Chinese-speaking financially motivated operator used it. Roughly 66,000 to 67,000 people and about 2,200 companies were affected. Hunt.io later found an exposed Hong Kong jump host whose state file showed Claude Code 2.1.221 started 510 times since 4 August 2026, with requests routed through a Claude Code Router and reseller domain to DeepSeek, plus GLM and Grok tooling, and 19 custom skills used 147 times for reconnaissance, web testing, and related tasks. The ARTEX GitHub page and Docker image were taken down on 11 October 2026, though copies remain online.

What is ThisBash / the Claude Dispatch flaw?

Forever Security researcher Gal Weizman named the technique ThisBash and published on 4 November 2026. Claude Dispatch's interface and API were reachable from an ordinary logged-in claude.ai session, so a browser extension that can control claude.ai could send commands to Claude Desktop as if they came from the user. The proof of concept used four prompts plus auto-clicked Allow buttons to install an npm package that opened Calculator. Because Dispatch spans devices, the desktop can be a different machine on the same Claude account, including a work laptop reached from a personal browser. Anthropic acknowledged the report as valid on 19 May 2026, closed it as a duplicate, and said Dispatch would be retired targeting end of October 2026; as of late October the support page still listed limited beta for existing users. Weizman notes the same trust pattern appears in Claude Code Remote Control and Cowork sessions.

Is GPT-6 Cyber a threat or a defensive tool?

Both, depending on who holds access. OpenAI's 5 November 2026 announcement presents GPT-6 Cyber as a cyber-specialized model for authorized vulnerability discovery and exploit development, initially for approved Daybreak Red customers via API and Codex. OpenAI rates it Critical for cybersecurity under its Preparedness Framework (up from High for GPT-5.6 Cyber). Reported research results include 81 novel high-severity Firefox vulnerabilities with Mozilla fixes underway, and 948 working Linux local privilege-escalation exploits with 304 critical findings disclosed to maintainers. Alignment evaluations claim 99.93% in-scope behavior without deployment prompts and 100% with them. For enterprises, the governance question is the same as for any powerful coding agent: who is allowed to run it, on which endpoints, with which secrets, and under what allow/deny policy.

What is SSHole and who is affected?

SSHole is a long-standing OpenSSH behavior on musl Linux (Alpine, OpenWrt, Buildroot, Yocto, and many Docker images) when password authentication is on and PAM is not used. Accounts whose shadow password field is only * accept any password between 257 and 1024 characters if another account on the host has a real password hash. Alpine's official AWS and Azure images ship a default alpine user with passwordless doas, so that path is effectively root. glibc distros, PAM setups, Dropbear, and OpenBSD are not affected the same way. OpenSSH 10.6p2 fixes it; turning PasswordAuthentication off is the interim control. The researcher who published sshole.com, Ozgun Kultekin, reported finding it with agent swarms.

How does Anomity help with this week's threats?

Anomity inventories eight AI artifact types on every managed endpoint - agents, MCP servers, extensions, plugins, skills, secrets, hooks, and CLIs - so Claude Code installs, custom skills, browser extensions that can reach AI surfaces, and plaintext API keys become visible before an attacker or a stolen-key mixer finds them. Secrets are redacted on the endpoint; only metadata leaves. On agents that expose a hook such as Claude Code PreToolUse, Anomity can allow, deny, or log a tool call before it runs, which is the control that matters when an agent is steered by a malicious extension, a poisoned skill, or an unsanctioned CLI. Findings route to SIEM, Slack, email, and Jira, and every change lands in a queryable 90-day audit trail. Anomity complements Network, EDR, DLP, and GRC; it covers the agent layer those tools were not built to see.

What should security teams do this week?

Patch OpenSSH to 10.6p2 on musl hosts or disable password authentication. Inventory AI agents, CLIs, skills, MCP servers, and browser extensions across the fleet, and flag personal AI accounts with cross-device remote control on corporate endpoints. Rotate LLM API keys and cloud credentials that may have been exposed in plaintext config or in third-party breaches. Restrict Microsoft 365 device-code authentication where unused, and treat unexpected Claude or ChatGPT installer links as hostile. For Daybreak Red or similar cyber-capable models, treat them as high-privilege agents: SSO, IP allowlisting, sandboxing, and human review for risky work. Book a demo if you want that inventory and hook-level governance in one place.

Ask AI about Anomity
ChatGPT Claude Perplexity Google AI Grok