Wednesday 2 September
importantWednesday, September 2, 20263 findings, 6 signals, 9 noted6 min read

A no-interaction iMessage chain installed Pegasus, multiple coding agents ran repository-selected Git helpers before trust, and PaperCut’s second emergency patch left attack paths open in the wild.

Pegasus delivered device-wide surveillance; coding-agent remediation remains partial, and PaperCut says Release 2 must be replaced by Release 3.

Priority findings3
01
Medium
Mobile — zero-click
Confirmed

An iMessage with no user interaction installed Pegasus and exposed private messages, files, the microphone, and the camera.

The demonstrated access was tied to a Pegasus operator.

Affects

Apple iPhones running an iOS build vulnerable to the undisclosed iMessage chain

What it enables

Zero-click installation of Pegasus with access to private data, encrypted messages, microphone and camera

The operator selects an iMessage-reachable targetThe target receives exploit content without interactingThe undisclosed iMessage chain executes and installs PegasusPegasus accesses files, private and encrypted messages, and the microphone and camera
Why this matters

Forensic evidence confirmed that a no-interaction iMessage chain installed Pegasus, which could access files, private and encrypted messages, the microphone, and the camera.

Detail and 3 sources
Required access

A Pegasus operator able to address the victim through iMessage; the victim need not open or acknowledge anything

Affected versions

Fixed as of iOS 18.4.1; Citizen Lab did not disclose the lower bound, device model or exact vulnerable builds

Proof of concept

Demonstrated by the researcher

Fix removes the capability

Not stated

The operator only had to address the target through iMessage. Exploit content arrived without an open or acknowledgment, executed the undisclosed chain, and installed Pegasus.

Once installed, Pegasus could reach files, private and encrypted messages, the microphone, and the camera.

Evidence
Citizen Lab confirmed both the zero-click iMessage infection and Pegasus presence from forensic artifactsSHARE Foundation independently reported the confirmed infection and absence of user interaction
02
High
RCE — coding agents
Provisional
CVE-2026-72718

A repository received as files can make multiple coding agents run attacker code before trust or approval.

The malicious .git configuration must survive delivery; an ordinary clone does not carry it.

Affects

Claude Code, Goose, Hermes Agent, Qwen Code, Grok Build, Codex and Cursor command-line AI coding agents on developer workstations.

What it enables

Pre-trust arbitrary code execution as the developer

Attacker prepares a repository directory whose .git/config names an attacker-controlled helper through core.fsmonitor or another command-bearing Git settingThe directory reaches the victim through an archive, shared drive, synchronization folder or removable media, preserving .git/configThe victim opens the folder or requests a review in an affected coding agentThe agent automatically invokes git status, git diff or equivalent context-gathering before its trust prompt, authentication step or approval boundaryGit refreshes its index and executes the repository-selected helper outside the agent sandbox with the developer's privileges
Why this matters

The changed boundary is execution before the agent asks for trust, authentication, or command approval across multiple products.

Detail and 1 source
Required access

Deliver a repository directory whose .git directory and attacker-controlled helper survive, then induce the victim to open or review it with an affected coding agent; an ordinary git clone does not preserve the malicious local configuration

Affected versions

Qwen Code 0.22.3 confirmed affected on 2026-09-01 and unpatched at publication, Grok Build 1.0.13 confirmed affected on 2026-09-01 and unpatched at publication, Hermes Agent 0.21.0 confirmed affected on 2026-09-01 and unpatched at publication, Claude Code ultrareview path confirmed affected on 2.1.252 and unpatched at publication, Claude Code core.fsmonitor path confirmed on 2.1.193 and fixed by 2.1.196, Goose 1.41.0 affected and fixed in 1.44.0, Codex and Cursor historical versions were affected and patched before publication; the primary research did not state exact ranges

Proof of concept

Demonstrated by the researcher

Fix removes the capability

No

The attacker prepares a directory whose .git/config selects a helper through core.fsmonitor or another command-bearing Git setting. The directory must arrive through an archive, shared drive, synchronization folder, or removable media with both the local configuration and helper intact.

When an affected agent gathers context with git status, git diff, or an equivalent command, Git executes the selected helper outside the sandbox as the developer. Manifold’s September 1 retest found current unpatched paths still exposed, so received repositories should be treated as executable content before an agent opens them.

Evidence
Manifold Security's original research explains the delivery constraint, sink and agent-by-agent demonstrationsManifold retested every unpatched path against a current release on 2026-09-01
03
High
Edge — PaperCut
Confirmed
CVE-2026-81578

PaperCut’s second emergency patch left takeover paths open in the wild.

Unauthenticated HTTP access can cross administrative configuration into JDBC-backed operating-system execution.

Affects

PaperCut NG and PaperCut MF print-management application servers on Windows, Linux and macOS.

What it enables

Unauthenticated configuration takeover followed by code execution in the PaperCut server process despite installation of an earlier emergency patch

Attacker sends a direct-component request that displays a public page while invoking an administrative ConfigEditor or UserList component.The authorization check validates the displayed page rather than the invoked component, allowing unauthenticated configuration changes.Attacker configures external user lookup with a malicious JDBC driver, URL and SQL statement.A forged user lookup activates Derby or H2 class-loading behavior and launches an operating-system process.On Windows, observed and demonstrated execution reaches LocalSystem; on Linux, the public Metasploit test runs as the papercut service account.PaperCut now says Emergency Patch Release 3 closes additional attack vectors observed in the wild and must replace Release 2.PaperCut has not mapped the attack vectors surviving Release 2 to the known request path, the additional authentication bypass, or another primitive.
Research leadDiff Release 2 against Release 3, or obtain PaperCut’s mapping of each added control to the bypassed request path and affected build.
Why this matters

Release 2 was not merely superseded: PaperCut says Release 3 closes additional attack vectors observed in the wild and instructs Release 2 users to replace it.

Detail and 4 sources
Required access

Unauthenticated HTTP reachability to the PaperCut NG/MF Application Server

Affected versions

All PaperCut NG and MF versions were treated as potentially impacted, Emergency Patch Release 1 was publicly bypassed, Emergency Patch Release 2 is not considered fully protective by PaperCut and Rapid7, Release 3 mitigation builds: MF v26 build 76531, MF v25 build 76532, MF v24 build 76534, NG v26 build 76530, NG v25 build 76533, NG v24 build 76535, Versions 23 and earlier must upgrade to a supported branch, All PaperCut NG and MF versions were treated as potentially affected; Huntress observed exploitation on 24.1.5.71847 and 25.0.10.75465 and reproduced it on NG 25.0.11.75758., Emergency Patch Release 3 provides mitigating builds for major versions 24, 25 and 26; older branches may require isolation or upgrade.

Proof of concept

Public exploit code

Fix removes the capability

Not stated

A direct-component request can show a public page while invoking an administrative ConfigEditor or UserList component. Because authorization follows the displayed page, an unauthenticated requester can alter external-user configuration and steer JDBC class loading or SQL execution into an operating-system process.

The public Metasploit test reached LocalSystem on Windows and the PaperCut service account on Linux. Release 2 users are not fully protected; Release 3 is the replacement.

Evidence
PaperCut states Release 3 closes additional attack vectors observed in the wild and instructs Release 2 users to replace it.Rapid7 documents the original patch bypass, says Release 2 users are not fully protected, and publishes the request and JDBC execution chain.Rapid7's public Metasploit pull request demonstrates SYSTEM execution on Windows, papercut-user execution on Linux, and explicitly verifies Release 2 against its known Release-1 bypass.
Signals6
important · Firmware — SonicWall

Two exploited SMA1000 flaws make an unauthenticated command-execution chain plausible, but the observed endpoint is unresolved.

Affects

SonicWall SMA 1000 Series enterprise remote-access VPN appliances, including physical and virtual deployments.

An Internet user can reach the Work Place portal without credentials. Public analysis shows how the SSRF and alternate access path could reach the separate operating-system command-injection flaw.

Detail and 2 sources

SonicWall confirms exploitation of both vulnerabilities and identifies fixed releases. We do not know whether observed intrusions completed this exact chain or which UID would run the command.

Composition to watch
Reach the Internet-facing Work Place portal without credentials.Use the SSRF and alternate access path to reach administrative functionality.Trigger the separate operating-system command-injection path.Whether a real intrusion completed this exact unauthenticated chain, and the resulting execution UID.
Unverified compositionFind telemetry containing both CVEs in one intrusion or reproduce the full chain while returning identity output.
important · RCE — vm2

vm2 3.11.7 stopped one nested import, not host execution through allowed external modules.

Affects

vm2, an in-process Node.js library used by plugin systems, code runners, CI tools and AI applications to execute untrusted JavaScript.

The remaining path requires NodeVM with require.external enabled and an allowed root containing a helper loaded in the host context. Importing a helper that exposes child_process or equivalent authority returns a capability the outer sandbox does not constrain.

Detail and 2 sources

GitLab validated that alternate path on 3.11.7. The configuration is narrower than a default sandbox escape, but the upstream README still presents require.external with root './' and warns against relying on vm2 alone.

important · Research — Artifactory

A default phantom or blank Artifactory cluster join key can be used to forge a cluster JWT and obtain an administrator token.

Affects

JFrog Artifactory Self-Managed, an enterprise repository for software packages, binaries, containers, and AI models.

On affected self-managed installations, the default phantom or blank join-key condition permits a forged cluster JWT, a service administrator token, and then a platform administrator token. That identity can reach users, credentials, federated topology, and stored artifacts.

Detail and 4 sources
important · Firmware — Virtualizor

A BGP hijack turned trusted Virtualizor updates into persistent root compromise.

Affects

Virtualizor, Linux-based virtualization and hosting-control software installed on hypervisor-management servers.

The attacker announced a more-specific route, diverted vendor traffic, and obtained a valid TLS certificate while certificate validation followed the diversion. Virtualizor then accepted a modified update package without independently checking a package signature.

Detail and 1 source

The malicious update executed as root and established persistence through an SSH key, an unauthorized account, and a systemd service.

important · Zero-click — Thunderbird

Thunderbird fixed a MIME path that uses uninitialized memory, but receipt-only triggering remains unknown.

Affects

Mozilla Thunderbird, a desktop email and calendar client on Windows, macOS and Linux.

A crafted MIME body can enter the affected error path. Public material does not say whether background synchronization is enough, whether preview or opening is required, or whether the stale value controls a security-sensitive operation.

Detail and 3 sources

Mozilla shipped fixes across release and ESR channels on September 1. Until trigger timing and the stale value’s consumer are known, this is a memory-safety primitive rather than an established receipt-only compromise.

Composition to watch
Send a crafted email containing a MIME body that reaches the affected error path.Determine whether synchronization, preview, or explicit opening invokes the path.Identify whether the uninitialized value reaches a pointer, length, or another security-sensitive operation.Whether synchronization alone triggers the path and whether the stale value has an exploitable consumer.
Unverified compositionObtain the restricted testcase or run malformed MIME bodies under MemorySanitizer while testing synchronization, preview, and explicit-open paths separately.
important · Zero-click — libheif

A WordPress Author turned a HEIF upload into server-side command execution through libheif 1.23.2.

Affects

libheif, the HEIF/AVIF decoding library used by desktop and server image-processing applications, including ImageMagick-backed WordPress media processing.

WordPress automatically passed the uploaded image through Imagick, ImageMagick, and libheif. Unequal component bit depths then made the mixed-interleave decoder write two-byte samples into a one-byte allocation, producing a controlled heap overwrite.

Detail and 3 sources

The researcher demonstrated file disclosure and command execution on WordPress 7.1. The path depends on a specific codec stack, and libheif 1.23.3 followed the 1.23.2 security release with a fix one week later.

Also noted9
Edge — Nexus 9000
f_a4bd83a2
Silicon One services on TCP/43210 and TCP/43211 are reachable through the default Layer-3 VRF and execute crafted input as root.
Cisco tested an infrastructure-ACL workaround; use its interactive checker to map each deployed PID and NX-OS release.
Research Cisco Nexus 9000 Series Switches Silicon One Remote Code Execution Vulnerability - Cisco
Wi-Fi — Cisco phones
f_a4d2ad91
Continuous crafted HTTP traffic can exhaust memory and disable affected Cisco SIP phones until a manual reboot.
Exposure requires network access and the normally disabled Web Access feature to be enabled.
Vendor Cisco Desk Phone 9800 Series, IP Phone 7800 and 8800 Series, and Video Phone 8875 with SIP Software Denial of Service Vulnerability
Firmware — SAUTER
f_de236bb5
An unauthenticated firmware-exchange race can bypass validation and potentially run unauthorized controller code.
Default exposure, reliable stock-hardware exploitation, and execution privilege remain unknown.
Vendor SAUTER: modulo 6 and EY-modulo 5 Vulnerability in Firmware update mechanism allowing remote code execution · not original
Edge — Switchvox
f_fbd21bcd
Switchvox’s unauthenticated /pa endpoint turns PhoneIP SQL injection into PostgreSQL COPY TO PROGRAM command execution.
Horizon3 demonstrated a reverse shell and observed exploitation attempts across honeypots.
Research CVE-2026-9586: Sangoma Switchvox RCE | Horizon3
Wi-Fi — TOTOLINK T6
f_4d1ae39b
An unauthenticated LAN client can recover Wi-Fi secrets, change administration and network settings, and install a custom CGI module.
The management interface is ordinarily LAN-reachable, with WAN exposure possible when remote administration is enabled.
Code / PoC CVE-Vendor-Coordination/TOTOLINK/README.md at main · ShengWu00/CVE-Vendor-Coordination · GitHub
Physical — Unitree
f_499e94f3
Physical Ethernet access can reach a published WebRTC-to-DDS and bashrunner chain that executes as root on a G1 EDU robot.
No shipped fixed firmware version was identified.
Research UniBLEed: Root RCE on Any Unitree G1 Humanoid Robot
Bluetooth — Unitree
f_34f9c624
A nearby BLE client with the target’s bound account or device key can still reach root on a G1 EDU robot.
Unitree blocked the cheaper arbitrary-account route, but not the credential-bound BLE and firmware primitives.
Research UniBLEed: Root RCE on Any Unitree G1 Humanoid Robot
Wi-Fi — Zephyr
f_da1b84a9
On-link traffic can trigger SiWx917 packet ownership errors that corrupt Zephyr’s shared packet pool, hang Wi-Fi, or crash the device.
The upstream fix removes the erroneous packet unreference.
Code / PoC SiWx91x WiFi driver double-unref / use-after-free of caller-owned TX net_pkt · Advisory · zephyrproject-rtos/zephyr · GitHub
Zero-click — libheif
f_8a679b60
A crafted HEIF conversion can encode adjacent libheif heap bytes into returned RGB pixels.
No cited source identifies a network-facing consumer that both performs the conversion automatically and exposes its output to the attacker.
Code / PoC heap OOB read / info disclosure (Op_YCbCr420_to_RRGGBBaa) · Advisory · strukturag/libheif · GitHub
What was checked · 2 quiet
Boot chain & TPMQuiet

No new boot-chain capability was confirmed; incomplete access to the cited U-Boot fix history prevents full closure.

Privilege escalationQuiet

PaperCut, Nexus Silicon One, and Cleo Harmony each crossed an authentication or role boundary into service-account or root execution.

fullchain.sh follows the day’s disclosures from bug to shell — what each one enables, what it links to, and where the fix left the primitive in place. A CVE is an input, not an event. When nothing qualifies, the brief says so. Every morning.

Wednesday, September 2, 2026