SOPHOS AI THREAT BRIEF

Daily brief

What changed in AI-driven cyber risk today

GitHub’s 2 November advisory says DBHub before 0.22.6 can still run writes and DDL on MySQL and MariaDB while read-only mode is on. JFrog’s 1 November reports say a Beta9 workspace token on a worker before 0.1.756 can read and write that worker through the sandbox file API or a chroot-only image build, and a volume identifier can mount the worker root unless the gateway is at least 0.1.776 and the worker is at least 0.1.756. Prefab UI’s generative sandbox through 0.20.2 remains the critical case published the same day. A tampered Claude Code extension was on Cursor’s marketplace for part of 29 October, and Microsoft’s 30 October analysis says the loader can leave an SSH key and startup scripts after the extension is removed. Ivanti’s privilege-escalation bug was due 2 November. The file-upload bug is due 13 November.

Tuesday, 3 November 2026

Read today's stories

Today's stories

DBHub read-only mode can still change MySQL and MariaDB

GitHub published GHSA-j656-3hf2-fvjc on 2 November. It is rated medium, and the advisory has no CVE. DBHub is Bytebase’s database MCP server. In versions before 0.22.6, read-only mode on the MySQL and MariaDB connectors can still run writes and DDL, including DROP TABLE. The parser treats every -- as the start of a comment. MySQL and MariaDB start a -- comment only when the dashes are followed by whitespace, a control character, or the end of the line. Otherwise the dashes are two minus signs and the rest of the line is ordinary SQL. The read-only check splits the text and classifies each piece, then the connector executes the original string. Those connectors open the pool with multiple statements enabled. A write placed after -- with no following space can pass the check and still run. GitHub’s check against the project’s own parser accepts that string as one read. The same string, sent to MariaDB 11.4 with multiple statements enabled, runs the read and then drops the table. The advisory says PostgreSQL is not affected by this path: PostgreSQL treats -- as a comment regardless, and it executes each split statement on its own. The report describes the bug on main at commit e99c3aa9, package version 0.22.3.

Read-only mode is the control operators use so an assistant can query a database without changing it. GitHub says anyone who can influence the SQL the assistant sends, most realistically through prompt injection in content the assistant reads, can perform writes and destructive DDL against MySQL and MariaDB. The first patched release is 0.22.6, published 10 July. The new fact is the advisory, published 2 November. The 24 June fix commit is an ancestor of the 0.22.6 tag and of the current release, v1.5.1, published 3 November. That tag’s own notes are other bug fixes. Its MySQL and MariaDB comment scanner still requires whitespace after --.

What to do

If an agent can send SQL through DBHub to MySQL or MariaDB, check the installed @bytebase/dbhub version. Anything before 0.22.6 is in the affected range. Move to 0.22.6 or later. v1.5.1 is the current GitHub release. Until that upgrade, do not treat read-only mode as a guarantee that the assistant cannot change data. PostgreSQL is outside this bypass. If untrusted content could have steered the assistant while an older build was connected, review recent writes and dropped objects on those MySQL and MariaDB databases.

Sources: GitHub Advisory Database, GHSA-j656-3hf2-fvjc, published 2 November; Bytebase DBHub advisory, GHSA-j656-3hf2-fvjc; DBHub release v0.22.6, 10 July; DBHub release v1.5.1, 3 November.

A Beta9 workspace token could read the worker hosting the sandbox

JFrog Security Research published three critical Beta9 reports on 1 November, each scored 9.9 (CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H). NVD published CVE-2026-120307, CVE-2026-120306, and CVE-2026-120308 the same day. Beta9, from beam-cloud, is that project’s serverless platform for GPU inference, sandboxes, and background jobs. Natan Nehorai of JFrog is the named researcher. All three are fixed in pull request 1897, commit a28bf12, released 15 September. CVE-2026-120307 is the sandbox file API, fixed on workers before 0.1.756. The worker serves that API from outside the container. If a path starts with a mount destination, including the default upload mount at /tmp/.beta9, the remainder is joined in a way that cleans parent-directory segments after the prefix check, so a long enough run of those segments names any path on the worker. A path that matches no mount is opened on the worker’s own filesystem. Upload, download, stat, list, find-in-files, replace-in-files, and delete-file all use that helper. A tenant with a workspace token and one running sandbox can read, write, list, and delete files on the hosting worker as root, including files of other tenants on that machine. JFrog’s read-only proof downloaded the worker pod’s Kubernetes service-account token through the file API for a path that did not exist inside the sandbox. In the shipped remote-node manifests that service account is bound to cluster-admin.

CVE-2026-120306 is the image build, on the same worker range. Each shell command in a tenant image becomes a Dockerfile RUN executed with Buildah. The worker image sets BUILDAH_ISOLATION=chroot, and the invocation does not pass another isolation mode, so the step keeps the worker pod’s PID, network, IPC, and user namespaces, runs as uid 0, and holds CAP_DAC_OVERRIDE. /proc/1/root is the worker container’s filesystem and is writable. That filesystem holds runc, buildah, and runsc, the service-account token, and the host directories mounted into the pod. Worker pods are privileged. Clip v2 layered builds are the default, so one image build reaches this. CVE-2026-120308 is the volume mount. A tenant deploying a function, endpoint, or task queue can supply a volume identifier. Nothing checks that the identifier names a volume that workspace owns. Parent-directory segments collapse the host path to the worker’s filesystem root, and the worker bind-mounts it read-write unless the volume was marked read-only. JFrog confirmed that on gateway 0.1.769 and worker 0.1.745. The proof read files on the worker, and JFrog says it was not run against a live multi-tenant cluster. Gateway versions before 0.1.776 and worker versions before 0.1.756 are affected. Stubs saved earlier are checked again when they launch, so both components need the upgrade. Worker 0.1.756 and gateway 0.1.776 are the tags cut from the fix commit. This morning’s worker tag 0.1.823 and gateway tag 0.1.843, both published 3 November, still contain that commit. Those tags are later releases, not the security fix itself.

What to do

If you run Beta9, upgrade every worker to 0.1.756 or later and every gateway to 0.1.776 or later. Upgrade both. A new gateway does not fix an old worker’s file API or image build, and an old worker will still mount a volume path an old gateway sends. Confirm the running worker and gateway versions, not only the client package. Treat a workspace token on an unpatched worker as access to that worker’s files and, where the shipped remote-node manifest is unchanged, to a cluster-admin service account.

Sources: JFrog Security Research, JFSA-2026-001699276, CVE-2026-120307, 1 November; JFrog Security Research, JFSA-2026-001699275, CVE-2026-120306, 1 November; JFrog Security Research, JFSA-2026-001699277, CVE-2026-120308, 1 November; NVD, CVE-2026-120307; NVD, CVE-2026-120306; NVD, CVE-2026-120308; GitHub Advisory Database, GHSA-6h4h-qffc-3gfp; GitHub Advisory Database, GHSA-xvhg-75x6-gj86; GitHub Advisory Database, GHSA-92w2-g272-fr77; Beta9 worker 0.1.756, 15 September; Beta9 gateway 0.1.776, 15 September; Beta9 worker 0.1.823, 3 November; Beta9 gateway 0.1.843, 3 November.

Prefab UI’s generative sandbox could read and write the host

GitHub published GHSA-3g72-rrqc-f76c, CVE-2026-120338, on 1 November. It is critical, scored 10.0 (CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H). Prefab UI’s Generative UI sandbox runs caller-supplied Python inside Pyodide, in a Deno subprocess, and the feature is documented as isolation for untrusted code, including interface code written by a model. From prefab-ui 0.14.0 through 0.20.2 the launcher starts Deno with unrestricted --allow-read, --allow-write, and --allow-net, and the runner leaves the Deno API on the JavaScript global object. Pyodide’s js module proxies that global, so the Python can call Deno.readTextFileSync and Deno.writeTextFileSync as the server’s operating-system user. JFrog Security Research, which published the same finding on 1 November, says the runner comment and the public docs describe the process as read-only. The usual remote path is FastMCP’s GenerativeUI provider. It registers the tool generate_prefab_ui and forwards that tool’s code argument into the sandbox. With no authentication on the server, any client who can reach it can call the tool. With authentication, any client allowed to call tools can. Generative UI is opt-in, and the host must have Deno installed. JFrog scores the authenticated case 9.9 (CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:H/A:H).

The sandbox is the control that was supposed to keep model-written code off the host. A caller who can submit the code can read and write files as the server user, and can plant a .pth file in the server virtualenv’s site-packages so a later Python process runs the planted import. JFrog says servers that never execute untrusted code in this sandbox are not exposed. The fix is already in a release that predates the advisory. GitHub and JFrog say 0.21.0 revokes Deno read, write, and network permission after startup, before it accepts code, starts the process with --no-prompt, and limits bootstrap network access to cdn.jsdelivr.net. The change is pull request 538, commit 030f074. Prefab tagged v0.21.0 on 26 October, and that tag’s changelog includes the sandbox pull request. A server left on 0.20.2 after that tag is still in the affected range.

What to do

If a host runs prefab-ui with Generative UI and Deno on the path, move to 0.21.0 or later and confirm the installed version. Until that upgrade, do not register GenerativeUI on a server untrusted clients can reach, which is JFrog’s interim mitigation. FastMCP servers that register the provider pick up the fix by running the patched prefab-ui release. Treat a reachable unauthenticated generate_prefab_ui tool on 0.14.0 through 0.20.2 as code execution as the server user, and review the virtualenv for an unexpected .pth file if that tool was exposed.

Sources: GitHub Advisory Database, GHSA-3g72-rrqc-f76c, published 1 November; JFrog Security Research, JFSA-2026-001699278, 1 November; Prefab release v0.21.0, 26 October.

A tampered Claude Code extension left an SSH key on Cursor machines

On 29 October Cursor reported that tampered copies of the Claude Code extension, anthropic.claude-code versions 2.1.312 and 2.1.313, were available in the Cursor marketplace from about 2:27 a.m. to 1:35 p.m. Pacific time. Cursor says Anthropic did not develop those builds, and that it has no evidence other versions of the extension or the official distribution were affected. The check Cursor publishes is for macOS and Windows installs from that window. The extension tried to install a persistent SSH backdoor: collect host and user details, append an attacker-controlled SSH key, enable the local SSH service, and open a reverse tunnel. Cursor says errors in the malicious code, and the need for administrator rights to turn on local SSH, blocked login through the tunnel on most macOS and Windows hosts. Cursor also says the extension still added the SSH key, installed persistence, and sent host details, and that the attacker server has since been taken offline. Cursor removed the versions, shut the marketplace down, rebuilt it, and updated the indicators on 30 October at 11:30 a.m. Pacific time.

Microsoft’s 30 October analysis of the injected JavaScript matches that chain and adds the point that matters after the marketplace takedown. The loader runs when the extension module loads. It looks for the comment svc-telemetry in the user’s authorized_keys file and appends an Ed25519 public key if the marker is missing. It then tries a reverse SSH forward to local port 22, writes startup scripts outside the extension directory, and later sends host details plus the names of environment variables whose names contain terms such as KEY, TOKEN, SECRET, PASS, AWS, or GCP. Microsoft says the callback sends those names, and that successful remote access depends on host privileges, SSH configuration, and the attacker’s server. The loader can still write the key and the persistence files when the tunnel fails. Microsoft confirmed the packages were never published on the Visual Studio Marketplace or as npm @anthropic-ai/claude-code, and Anthropic confirmed to Microsoft that Anthropic’s systems were not compromised. Removing the extension leaves the key, the scripts, and the startup entries in place.

What to do

If anyone installed or updated anthropic.claude-code 2.1.312 or 2.1.313 from the Cursor marketplace on 29 October, search authorized_keys for the comment svc-telemetry and the fingerprint SHA256:0Z6R+DF64Do2nwQic8a7kkfDLNNk4IiK+2U2Rci4htk. On macOS look for ~/Library/LaunchAgents/com.apple.sysmond.plist. On Linux look for a user crontab @reboot entry. On Windows look for a scheduled task named SysMon and a Startup-folder sysmon.cmd. On each platform also look for ~/.config/.smt and ~/.config/.smtk, or the Windows equivalents under the user profile. Preserve the extension package, the authorized-keys file, and those startup objects, then remove the confirmed key and persistence. Cursor’s remediation scripts are linked from its notice. Rotate credentials and sessions from a clean machine when the host ran the loader. Hunt for connections to 134.209.82.225, the address in Microsoft’s hunting query and Cursor’s indicator list. A missing extension, and the offline attacker server, both leave those artifacts behind.

Sources: Cursor security notice, Claude Code extension, 29 October, updated 30 October; Microsoft Security Blog, malicious Claude Code extension, 30 October.

Ivanti Connect Secure privilege-escalation deadline was 2 November

On 30 October CISA added CVE-2026-111357 and CVE-2026-118527 to its Known Exploited Vulnerabilities catalog and revised its 13 October Ivanti alert the same day. CVE-2026-111357 is a missing authentication check for a critical function that CISA says lets a local authenticated attacker escalate privileges. The federal due date is 2 November, forensic triage is required, and ransomware use is listed as unknown. CVE-2026-118527 lets a remote authenticated attacker upload files with attacker-controlled content to a known location on disk. Its federal due date is 13 November, CISA lists forensic triage required as No, and ransomware use is unknown. CISA says each can be chained with the other and with CVE-2026-104935, the authentication-bypass bug added on 13 October, whose federal due date was 16 October and which also requires forensic triage. That older bug allows a remote unauthenticated attacker to execute arbitrary commands. CISA’s revised alert says the three affect Ivanti Connect Secure 25.1.1.2 and prior, 25.1.2.2 and prior, and 22.8R2.5 and prior, and that actors are using them for remote unauthenticated command execution, credential theft, and data exfiltration. Version 22.8 reaches end of life on 31 January 2027.

The federal due date for the privilege-escalation bug was 2 November. The CISA catalog feed fetched for this edition is still version 2026.10.30, released 30 October, with 1,766 entries. The unauthenticated command execution has been on the exploited list since 13 October. CISA says Ivanti’s Integrity Checker Tool may miss CVE-2026-104935 because the commands run in memory. A gateway that only cleared that check, or that patched only the October bug, can still be on a release CISA lists as affected. Forensic triage is required for CVE-2026-111357 and CVE-2026-104935, so a reachable appliance is a possible compromise as well as a missing update.

What to do

Find every Connect Secure appliance on 25.1.1.2 or earlier, 25.1.2.2 or earlier, or 22.8R2.5 or earlier, and update it to the latest version in Ivanti’s advisory. Treat CVE-2026-111357 as overdue. Its federal due date was 2 November. CVE-2026-118527 is due 13 November. Where compromise is suspected, follow the recovery sequence CISA quotes from Ivanti: back up the appliance configuration, factory-reset and upgrade (rebuild cloud and virtual instances, then upgrade), restore that configuration, reset passwords for domain, administrator, and local accounts, revoke and reissue stored certificates, and reset stored API keys and license server credentials. Preserve volatile evidence before you change the system when you can. Treat a clean Integrity Checker Tool result as incomplete for CVE-2026-104935.

Sources: CISA alert, two exploited vulnerabilities, 30 October; CISA alert, Ivanti Connect Secure, revised 30 October; CISA Known Exploited Vulnerabilities catalog, CVE-2026-111357 due 2 November, CVE-2026-118527 due 13 November; CISA, BOD 26-04 forensic triage guidance.