A researcher writing as m00dy published findings on 7 October 2026, coordinated through CERT/CC (VU#765030), and CVE-2026-93034 was published on 8 October for SGLang, a widely used serving engine for large language models. SGLang's processes pass internal messages over ZeroMQ. By default they use Python Pickle (SGLANG_USE_PICKLE_IPC=true), which can run code while loading data; operators can switch to MessagePack, but the MessagePack path still accepts a PickleWrapper record and immediately unpickles it without a type allowlist or authentication. On a stock v0.5.18 service running a real model across two GPUs, an unauthenticated peer on a second machine got its code to run inside the detokenizer process, with the setting both on and off. In the default single-node layout the receiver is local to the host; it becomes reachable over the network when data-parallel attention is enabled with a non-loopback --dist-init-addr. Source inspection found the same code in v0.5.20 and main on 1 October; the researcher had not assessed v0.5.21 (released 2 October), so no fixed version is confirmed. CVSS 3.1 9.8 (CERT/CC). No exploitation reported. Primary: researcher write-up and CVE record.
m00dy — Disabling Pickle did not remove Pickle (SGLang, CVE-2026-93034 / VU#765030; 7 Oct 2026)
IBM published a security bulletin on 7 October 2026 for IBM Security Verify Access 10.0 to 10.0.9.2 and IBM Verify Identity Access 11.0 to 11.0.3, the access-management and reverse-proxy gateway (formerly ISAM/WebSEAL) many organisations put in front of web applications for single sign-on and multi-factor login. The worst issues, CVE-2026-78401 and CVE-2026-78406 (both CVSS 9.8), let a remote attacker with no login run arbitrary code through deserialisation of untrusted data (CWE-502). The bulletin lists 22 CVEs in all. It also covers CVE-2026-16916 (9.1, authenticated code execution through a protection mechanism failure in the appliance), CVE-2026-17189 (8.2, reflected cross-site scripting), CVE-2026-19878 (6.5, bypass of extra authentication workflow steps) and denial-of-service, log-forgery and local management interface flaws. Appliance and container editions are both affected. IBM lists no workarounds and no exploitation. Primary: IBM.
IBM — Security Bulletin: vulnerabilities addressed in IBM Verify Identity Access and IBM Security Verify Access (7 Oct 2026)
CVE-2026-107181 was published on 7 October 2026 for a flaw in Telegram Desktop before 7.2.9 that researcher beaksec described in a technical write-up first published on 3 October and updated on 7 October. When a link is opened from outside Telegram (for example from a browser), a second instance passes it to the running app over a local inter-process channel that did not escape its record separator (CWE-143). A crafted tg:// link with an unescaped semicolon could therefore inject an extra command that reached an old internal helper, once used for release publishing, which uploads a local file to a chat without any permission check or prompt. Files that can be taken include Telegram's tdata session keys, so an attacker can take over the account when no local passcode is set. The demonstrated chain was on Windows with version 6.9.3 and is said to work through 7.2.8; it also relies on group files downloading automatically and on anyone being able to add the victim to a group, and the browser may ask before launching the app. Links clicked inside Telegram are not affected. Telegram fixed it in commit db3405699f (16 September) and released 7.2.9 on 17 September 2026 without a separate advisory; CVSS 3.1 is 8.1. No exploitation in the wild is known. Primary: CVE record (NVD) and Telegram's fix; wire: ThreatWire.
German security firm Jakkaru disclosed on 7 October 2026 a business-logic flaw in iSolarCloud, the cloud platform Sungrow uses to manage solar plants, inverters and battery storage. One value of the login_type field in the (encrypted and signed) login request made the service authenticate the account named in the request and ignore the password, so anyone who knew a user's email address could log in, with no email or other login alert to the owner, and could then use account recovery to keep control. Customer and administrative accounts share the same management environment, so Jakkaru says a compromised administrator account could view and change plants, start or stop inverters and batteries, see registered organisations and users (including large German solar installers), and install custom firmware on any cloud-connected device. iSolarCloud runs four regional systems, European, Chinese, Australian and international, and the flaw was in all of them. Jakkaru says Sungrow published a hotfix within a day of being told and began a root-cause review. CVE-2026-107194 was published on 7 October, and no exploitation is reported. Sungrow is one of the largest inverter makers, with more than 1,000 GW of converters installed by Jakkaru's count, and is common on Australian rooftops. Primary: Jakkaru; wire: Cyber Security News. Correction (22 Oct 2026): this card said no CVE had been published. CVE-2026-107194 was published on 7 October, and Sungrow's own advisory SGSA-202608-0028 (10 October) rates it CVSS 4.0 9.2 Critical, says the flaw was fully remediated on 25 August 2026 with no customer action needed, and describes the access an attacker gained as viewing the account's device list, operating status and historical monitoring data; the wider impact on this card is Jakkaru's assessment. (source: Sungrow security advisory SGSA-202608-0028)
Jakkaru — Sungrow vulnerability exposes gigawatts of power worldwide (7 Oct 2026)
curl 8.23.0 was released on 14 October 2026 with 24 security fixes. The standout is CVE-2026-92392, HTTP/2 upload Use-After-Free, rated HIGH by the curl project (CWE-416). During an HTTP/2 upload, mishandling of a server RST_STREAM (HTTP_1_1_REQUIRED or REFUSED_STREAM) can leave a dangling pointer to a Curl_easy handle; when the application frees that handle and later cleans up the multi handle, libcurl can download attacker-controlled data into the freed memory and then call a corrupted function pointer — remote code execution. Affected: curl 8.12.0 through 8.22.0; not affected: before 8.12.0, 8.23.0 and later, plus the older-branch fixes 8.14.3, 8.16.2 and 8.20.2 listed in the advisory. Daniel Stenberg says successful RCE needs many concurrent conditions (valid server certificate for an upload of roughly the right size, known architecture/libc/heap layout, nghttp2 and OpenSSL/TLS 1.3 allocation behaviour, ~23-byte payload window), so wild exploitation is judged extremely difficult and a crash is the likelier outcome — hence HIGH not CRITICAL. The bug was found and reported by Luna Tong at Anthropic using Anthropic's OSS Scanner on 15 September; the release was brought forward by two weeks. Primary: curl CVE-2026-92392 advisory and Daniel Stenberg's UAF write-up; release notes for 8.23.0. UPDATE 15 October: the reporter's HackerOne write-up (report 4030333), with the malicious server script and a Docker container described as reproducing the code execution, became public alongside the advisory on 14 October, so a proof of concept is available; no source reports attacks using it. curl's machine-readable advisory lists the fixed versions as 8.14.3, 8.16.2, 8.20.2 and 8.23.0. Debian 13 (trixie) is fixed in 8.14.1-2+deb13u7 (DSA-6558-1); Debian 12 (bookworm) does not carry the vulnerable code. The Hacker News says the attack needs the client to upload a file of roughly 135 KB over HTTP/2 to an attacker-controlled server whose certificate passes the client's checks, and that libcurl negotiates HTTP/2 by default for https:// URLs when built with HTTP/2 support. curl's interim advice is not to use HTTP/2. Wire: The Hacker News.
curl — CVE-2026-92392: HTTP/2 upload Use-After-Free (14 Oct 2026)
Splunk published five security advisories on 7 October 2026. SVD-2026-1001 lists 17 CVEs in Splunk Enterprise; the worst, CVE-2026-76268 (CVSS 3.1 9.8), affects 10.4 before 10.4.3 and 10.2 before 10.2.7: an unauthenticated attacker with network access to the Patroni REST API on a search head cluster member can run operating-system commands, because that interface does not require authentication for critical configuration operations. Versions 10.0.x and 9.4.x are not affected by that one. The same advisory also fixes a local privilege escalation to root through Linux package upgrades (CVE-2026-76266, 7.7), low-privilege access to other users' search jobs, queries and results (CVE-2026-76269, 6.5), SQL injection in the SPL2 module catalog (CVE-2026-76270, 6.5), an SSRF in the Splunk Observability Cloud app (CVE-2026-76274, 6.5) and several Splunk Secure Gateway authorisation gaps (CVE-2026-76265, 76272, 76280). SVD-2026-1002 groups internally found issues into five CVEs, the highest scored 9.8 (CVE-2026-76281, improper access control) and 9.0 (CVE-2026-76284, improper neutralisation). SVD-2026-1003 and SVD-2026-1005 cover third-party package updates in Splunk Enterprise and the Splunk Add-on for AWS, both rated Critical. SVD-2026-1004 fixes Splunk MCP Server before 1.2.1, which could send a user's Splunk authentication token to the URL of a custom API tool set up by another user (CVE-2026-76286, 5.3). Splunk does not report exploitation. Primary: Splunk security advisories.
Splunk — SVD-2026-1001: Security Vulnerabilities in Splunk Enterprise, September/October 2026 (7 Oct 2026)
HashiCorp published four Vault security bulletins on 7 October 2026, all fixed in Vault Community Edition 2.1.2 and Vault Enterprise 2.1.2, 1.21.12, 1.20.17 and 1.19.23. HCSEC-2026-43 (CVE-2026-89322, CVSS 3.1 7.2): Vault did not always evaluate ACL policies against the canonical form of resource and policy names, so an authenticated user with delegated permissions could get around an explicit deny, reach a protected resource or assign a denied policy, and so escalate privileges. HCSEC-2026-41 (CVE-2026-105816, 8.0): with Shamir seals and an external plugin directory configured, a privileged operator who can restore an Integrated Storage (Raft) snapshot may be able to run arbitrary code on the Vault host, because stored plugin catalog entries were not checked to point inside the plugin directory. HCSEC-2026-40 (CVE-2026-105818, 5.9): under the default directory policy, the PKI secrets engine's ACME server could issue certificates containing identities the ACME challenge never validated, which could allow impersonation toward systems that trust that PKI mount (affects 1.14.0 and later). HCSEC-2026-42 (CVE-2026-105820, 5.4, Enterprise only): crafted policy names with path traversal let a token use policies from other namespaces, including the root namespace. HashiCorp does not report exploitation. Primary: HashiCorp security bulletins and CVE records.
HashiCorp — HCSEC-2026-43: Vault inconsistent ACL policy evaluation may allow bypass of deny restrictions (7 Oct 2026)
vulnerabilitiesidentitycloud
Vulnerabilities
Wed 7 Oct
CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H (NGOAM, MPLS OAM AND NX-API ON NEXUS)
Cisco published its scheduled October 2026 batch at 16:00 UTC on 7 October (midnight 8 October Perth). The headline items are three advisories for Nexus 3000 and Nexus 9000 switches running standalone NX-OS. Next Generation OAM (NGOAM), the VXLAN and SRv6 diagnostics feature, carries three flaws (CVE-2026-76485, CVE-2026-76486, CVE-2026-76501) where crafted IP packets sent to any IP interface let a remote attacker with no login run code as root or crash and reload the switch. MPLS OAM has the same outcome from a crafted MPLS echo-request (CVE-2026-76465). The NX-API management interface has a third path, a crafted HTTP request (CVE-2026-76471). All five score 9.8. Exposure depends on features: CVE-2026-76485 needs only NGOAM enabled, CVE-2026-76486 also needs SRv6 or NV Overlay with a VXLAN peer learned, CVE-2026-76501 also needs SRv6, and MPLS OAM and NX-API are both off by default. On UCS 6300 fabric interconnects the NX-API bug is reachable through the UCS Manager XML API with a low-privileged login, so Cisco rates it High there. In the same release Cisco grouped internally found bugs into Critical 'security hardening' advisories: NX-OS (six CVEs, top score 9.8 for CVE-2026-76455), APIC, the controller for Nexus 9000 in ACI mode (CVE-2026-76498, -76499 and -76500, each 9.8), and Meraki (seven CVEs, top 9.6). It also fixed a Finesse SSRF (CVE-2026-20362, 7.2) that Cisco says has been publicly described, and medium-rated NX-OS, APIC and ACI bugs. Cisco's own engineers found the switch flaws, and PSIRT knows of no public disclosure or malicious use of them. For Australian operators, these are core data-centre and campus switches, so treat the OAM and NX-API advisories as the priority on any switch where those features are on.
Cisco Security Advisory — Nexus 3000 and 9000 Series Switches NGOAM Remote Code Execution Vulnerabilities, cisco-sa-ngoam-rce-LWKQ4BU (7 Oct 2026)
Cisco published two Critical advisories on 7 October 2026 for Cisco License On-Prem, the on-premises licensing server previously sold as Smart Software Manager (SSM) On-Prem. The first covers four bugs. CVE-2026-20328 (CVSS 9.1) is a broken check in the password-reset process of the web management interface: a remote attacker with no account can reset the password of any user, including administrators, and log in as them. CVE-2026-76454 (CVSS 9.1) is in the Smart Licensing Utility API, which lacks authentication and input validation, so a crafted request can write arbitrary files on the server or knock it over. Two lower-rated bugs, an authenticated OS command injection through submitted configuration (CVE-2026-76437) and an authenticated SQL injection (CVE-2026-76452), score 4.9 each. Gabriele Paris of the NATO Cyber Security Centre reported CVE-2026-20328, -76437 and -76452, and Trung Nguyen of CyStack reported CVE-2026-76454. The second advisory is a Critical 'security hardening release' grouping four internally found bugs (CVE-2026-76480 at 9.8, CVE-2026-76482 at 10.0, CVE-2026-76483 at 9.1, CVE-2026-76484 at 8.8). Both advisories say every configuration is affected and there are no workarounds. Cisco PSIRT knows of no public disclosure or malicious use. License On-Prem servers usually sit inside the network holding entitlement data for every Cisco device an organisation licenses, so an exposed management page is the main worry.
JFrog disclosed CVE-2026-105192 on 7 October 2026, a critical flaw in LMCache, the open-source key-value cache that speeds up LLM serving engines such as vLLM. In multiprocess (distributed) mode LMCache runs as a standalone server and opens a ZeroMQ socket, port 5555 by default, so worker processes can register and share cached blocks. That socket has no authentication. One message type carries a msgpack extension that is handed to Python's pickle.loads while the server is still decoding the REGISTER_KV_CACHE arguments, before any handler or type check runs, so a single crafted message executes the sender's code as the LMCache user. JFrog says the official container images run that process as root. The 9.8 score applies when the server is bound to a routable address with --host, which is the documented multi-node setup and what LMCache's own example Kubernetes DaemonSet does (it listens on every interface); a default single-host install that stays on localhost is not reachable from other machines, and LMCache running inside a single vLLM process does not open the port at all. The decode path arrived in 0.3.9 (October 2025) and is still present in 0.5.5, the latest PyPI release, the 0.5.6 release candidates through rc3 and the dev branch as of 7 October. LMCache has not published its own advisory, and JFrog gives no way to tell whether a server has already been hit. Separately, one GitHub account opened six more LMCache security reports on 6 October alleging unauthenticated cross-tenant cache access and command-running network services; those have no CVE or maintainer confirmation. Researcher: Yuval Moravchick (JFrog). Primary: JFrog Security Research advisory JFSA-2026-001694382; wire: The Hacker News (7 Oct). Correction (22 Oct 2026): this card said no fixed release existed and that LMCache had not published its own advisory. LMCache published advisory GHSA-w5f2-9q9c-ff4h on 13 October (CVSS 3.1 9.8) listing 0.3.9 up to 0.5.6 as affected and 0.5.6 as the fix; 0.5.6 was released on PyPI on 15 October. (source: LMCache GitHub advisory GHSA-w5f2-9q9c-ff4h)
JFrog Security Research — LMCache unauthenticated RCE via pickle deserialization on the multiprocess ZMQ transport, CVE-2026-105192 (7 Oct 2026)
The Harbor project published GitHub security advisory GHSA-2phj-cp9f-qq6w on 6 October 2026 (no CVE assigned): a project administrator can point an HTTP or Slack webhook at localhost, private ranges, link-local, or the cloud metadata endpoint 169.254.169.254. When an event fires, Harbor jobservice POSTs from inside the deployment; if the target returns an error status, the response body is written to the webhook execution log readable by project members — so a ProjectAdmin can reach internal services and skim error responses (including metadata-style leaks). CVSS 3.1 8.5 (AV:N/AC:L/PR:L/UI:N/S:C/C:H/I:L/A:N). Affected: Harbor with webhooks from 1.7.0 onward when untrusted users can become project admin (default Project Creation: Everyone makes every registered user a project admin of projects they create) and jobservice can reach internals. Fixed in 2.13.6, 2.14.5, 2.15.3 and 2.16.0+, which reject private/loopback/link-local webhook targets unless HARBOR_ALLOW_PRIVATE_NETWORK_ACCESS / network.allow_private_network_access is deliberately re-enabled. Until then: set Project Creation to Admin Only, restrict jobservice egress, block 169.254.169.254, prefer IMDSv2. Wire: The Hacker News ThreatsDay (16 Oct) via OX Security. Primary: Harbor GHSA.
Harbor GHSA-2phj-cp9f-qq6w — SSRF in webhook delivery (6 Oct 2026)
wolfSSL released wolfSSH 1.6.0 on 6 October 2026 with five security fixes. CVE-2026-16516 (Critical per wolfSSL): the client did not check that the ECDSA curve in a server's host key matched the negotiated algorithm, so a man-in-the-middle could swap in a key on another curve and pass signature verification with its own private key; it also needs the application to use a lax public-key check callback. Affects all versions through 1.5.0. CVE-2026-83540 (High): wolfSSHd on Windows shared one authentication context and its logon token across concurrent connections, so a user with a valid account could end up logged in as another, more privileged user (1.4.15 to 1.5.0; non-Windows builds unaffected). Medium: CVE-2026-84897 lets an unauthenticated client make a server run client-side DH group-exchange handling and costly primality checks (about half a second of CPU per 1 KB packet); CVE-2026-81535 admits forwarded-tcpip channels without the forwarding policy callback in --enable-fwd builds; CVE-2026-83742 is an SFTP path bug in wolfSSH_RealPath() that writes one NUL byte past a stack buffer and can crash the process. 1.6.0 also turns on strict key exchange (Terrapin protection) by default, requires RSA user keys of at least 2048 bits and limits failed logins to six. wolfSSH is mostly found embedded in devices and firmware, so fixes arrive through vendor updates. No exploitation is reported. Primary: wolfSSH GitHub release notes; wire: Cyber Security News.
wolfSSL — wolfSSH v1.6.0-stable release notes (6 Oct 2026)
