Security news, decoded.

Researchers report that Yarbo Android and iOS apps contain hard-coded MQTT credentials, tracked as CVE-2026-10557, which can be extracted from the apps and used to access MQTT brokers supporting a large global robot fleet. A second issue, CVE-2026-7368, is a lack of proper per-device/per-user authorization in the Yarbo cloud, meaning a single valid login could allow fleet-wide access. This matters because attackers may be able to subscribe to telemetry and issue commands across many robots rather than affecting only one device.

NVIDIA has released an urgent security update for its NeMo Framework, addressing three High-severity code-injection-related flaws: CVE-2026-24155, CVE-2026-24252, and CVE-2026-24228. The issues can enable remote or local code execution depending on the CVE, including potential privilege escalation and data tampering, and affect all NeMo versions from 0.0 up to 2.7.2. Users should upgrade to version 2.7.3 or later to reduce the risk of exploitation, especially on shared training or AI pipeline infrastructure.

MongoDB Server has disclosed a new set of security issues affecting document databases, with fixes covering several MongoDB release lines (including 7.0, 8.0, 8.2, and 8.3). The advisories include CVE-2026-11933 (use-after-free in server-side JavaScript that can leak memory or crash when reachable by an authenticated user), CVE-2026-9740 (an unauthenticated denial-of-service crash via BSON validation recursion), CVE-2026-9750 and CVE-2026-9743 (authenticated and aggregation-related crash or incorrect-results scenarios). Because at least one bug can be triggered without authentication, upgrading to the provided fixed versions such as 8.0.26, 8.2.11, and 8.3.4 is a priority.

Google Cloud Vertex AI SDK for Python contained a bug that let an attacker with their own Google Cloud project and the victim’s project ID hijack model uploads and get code to execute in Google’s model serving infrastructure. The issue is in the SDK’s default temporary Cloud Storage bucket handling, enabling cross-tenant model replacement (notably for pickle/joblib-based models) and theft of OAuth tokens from the serving container; Unit 42 reported the risk and Google patched it in v1.148.0 (update from versions such as 1.139.0 and 1.140.0). No CVE has been published for this specific problem as of the article’s publication, but users should upgrade and explicitly set stagingbucket to a bucket they control.

Researchers say a ShapedPlugin supply chain attack compromised the vendor’s distribution pipeline and inserted malicious backdoor code into premium Pro plugin releases, impacting WordPress sites that update normally. The activity is tracked under CVE-2026-10735 and CVE-2026-49777, with confirmed exposure including Real Testimonials Pro version 3.2.5 plus other premium plugins such as Product Slider Pro and Smart Post Pro. This matters because the malware also targets authentication (including 2FA/TOTP secrets), potentially allowing attackers to bypass multi-factor protections.

CISA has directed U.S. federal agencies to patch within three days a actively exploited vulnerability in the LiteSpeed cPanel user-end plugin, tracked as CVE-2026-54420. Reported as CVE-2026-48172, the high-severity flaw affects user-end plugin versions before 2.4.8 and can let attackers with FTP or web shell access escalate to root on shared hosting systems running CloudLinux/CageFS due to a UNIX symlink-following issue. This is included in CISA’s Known Exploited Vulnerabilities Catalog, making timely remediation critical to reduce the risk of widespread compromise.

Researchers report two new Windows builds of the SprySOCKS backdoor—tracked internally as WINDRV and WINPLUS—previously believed to target Linux. The Windows versions use hard-coded command-and-control with TCP, UDP, and WebSocket communications and expand capabilities for reconnaissance, service/process management, and file operations, with WINDRV leveraging kernel drivers and WINPLUS using the Windows Print Spooler to load the payload more covertly. Evidence suggests UEFI-boot persistence may be involved, potentially exploiting CVE-2023-24932, a Windows Boot Manager security bypass that Microsoft fixed in May 2023.

Threat actors are reportedly exploiting multiple critical Fortinet FortiSandbox vulnerabilities in the wild, enabling privilege escalation and remote code execution without authentication or user interaction. The affected flaws are tracked as CVE-2026-39813, CVE-2026-39808, and CVE-2026-25089, and Fortinet released security updates on April 14—organizations should upgrade to the latest fixed versions to stop active attacks. This matters because FortiSandbox weaknesses have previously been leveraged in ransomware and intrusion campaigns to break into target networks.

ESET found Windows versions of the SprySOCKS backdoor being used in attacks against government organizations in Taiwan, Thailand, Pakistan, and Honduras during 2023–2024. The activity is assessed with high confidence as the work of the Earth Lusca threat actor (also tracked as FishMonger, Aquatic Panda, Red Dev 10, and TAG-22), and the malware’s Windows variants add kernel-level stealth plus covert command-and-control through diverted TCP traffic. Researchers also noted signals that a possible UEFI bootkit component could involve CVE-2023-24932 (a Secure Boot issue), which matters because it may indicate additional pre-OS persistence risk alongside the already stealthy Windows backdoors.

The lead maintainer of the curl project announced a temporary halt in accepting vulnerability reports for the period July 1, 2026 (00:00 CEST) through August 3, 2026 (09:00 CEST), including submissions via HackerOne or email. The pause is intended as a scheduled rest rather than a response to a security incident, and it also shifts the planned release of version 8.22.0 to September 2, 2026. While urgent issues may still be handled earlier for companies with paid support contracts, other reports will be deferred, which matters for organizations relying on fast disclosure handling.

Attackers are using phishing emails that deliver obfuscated script-based loaders to trigger an advanced Agent Tesla infection chain, moving from attachment execution to process injection, stealthy data theft, and exfiltration with minimal user awareness. The article highlights how this modern approach shifts away from older Microsoft Office exploitation commonly associated with CVE-2017-11882, CVE-2017-0199, and CVE-2018-0802, while still referencing CVE-2023-24059 in the context of the observed threat. This matters because the reliance on in-memory payload execution, process hollowing, and anti-analysis checks can make traditional detection harder, increasing the risk of credential and session compromise.

Cisco has released security updates for CVE-2026-20262, a medium-severity vulnerability in the web UI of Cisco Catalyst SD-WAN Manager (formerly SD-WAN vManage) that is being actively exploited in the wild. The issue can let an authenticated remote attacker abuse a file upload path-handling weakness to create or overwrite files on the device’s filesystem, which may be leveraged toward higher privileges depending on attacker access. The fix is available across multiple Cisco Catalyst SD-WAN releases, including Cisco Catalyst SD-WAN Release 20.9.9.2, 20.12.7.2, 20.15.4.5, 20.15.5.3, 20.18.3.1, and 26.1.1.2, covering Cisco Catalyst SD-WAN Manager On-Prem, Cisco SD-WAN Cloud-Pro, Cisco SD-WAN Cloud (Cisco Managed), and Cisco SD-WAN for Government (FedRAMP).

Tornado maintainers have released version 6.5.6 to patch three security issues in the Python web framework, including CVE-2026-49853, CVE-2026-49855, and CVE-2026-49854. CVE-2026-49853 (7.7) can leak Authorization headers when SimpleAsyncHTTPClient follows redirects to a different origin, potentially exposing credentials across sites. CVE-2026-49855 (7.5) addresses a gzip bomb that could exhaust memory, while CVE-2026-49854 (5.3) fixes an out-of-bounds read in Tornado’s optional native extension that could reveal small amounts of uninitialized memory. Upgrading to Tornado 6.5.6 is the recommended mitigation, especially to address the credential-leak risk.

A critical issue in the Haskell library crypton-x509-validation (CVE-2026-9648, CVSS 9.1) allows TLS clients to accept forged certificates because the library does not enforce X.509 NameConstraints as required by RFC 5280. This matters because it can let attackers extend trust beyond the permitted scope of a name-constrained CA, potentially enabling credential/session interception in real-world deployments, especially delegated PKI used by financial institutions. CERT/CC reports that upgrading to crypton-x509-validation version 1.9.1 is the recommended mitigation.

phpBB has disclosed a critical authentication bypass, tracked as CVE-2026-48611, that can allow an unauthenticated attacker to log in as arbitrary users (including administrators) by abusing phpBB’s OAuth-related session handling. The flaw affects phpBB versions up to and including 3.3.16, and the 4.0.0 alpha branch, meaning many installations are exposed by default. This is a severe risk because it enables account takeover via a crafted request, and administrators should update to 3.3.17 (or apply the official mitigations/workarounds) as an urgent priority.

CVE-2026-54420 is a privilege escalation issue in the LiteSpeed cPanel plugin (before version 2.4.8) that is reportedly under active attack, allowing a low-privileged tenant on shared hosting to escape isolation and obtain full root access. The flaw involves improper handling of user-controlled symlinks when the server runs CloudLinux/CageFS, which matters because compromising one account can put all sites on the same host at risk. LiteSpeed has remediated the problem in cPanel plugin v2.4.8 (bundled with WHM Plugin v5.3.2.1), so administrators should patch immediately and investigate logs for suspicious activity.

Researchers describe a new Microsoft Copilot attack dubbed “SearchLeak” that can let an adversary exfiltrate user-accessible Microsoft 365 data, including emails, meeting notes, OneDrive files, and SharePoint documents, using a crafted Copilot Search link with no extra victim action. Varonis Threat Labs says the multi-stage issue relies on a lesser-known parameter-to-prompt Injection (P2P) pattern and can bypass protections by embedding attacker-controlled content in ways that trigger AI behavior before sanitization. Microsoft has patched the vulnerability as CVE-2026-42824, underscoring how prompt-injection weaknesses in LLM-based enterprise assistants can turn internal data access into an attacker-controlled output channel.

Researchers at Obsidian Security report a multi-step vulnerability chain in LiteLLM that can allow a default low-privilege account to escalate to full proxy admin and achieve code execution. The affected issues are tracked as CVE-2026-47101, CVE-2026-47102, and CVE-2026-40217; together they can bypass authorization, elevate privileges, and escape the Custom Code Guardrail’s sandbox. Because LiteLLM sits in the middle of AI requests, a takeover can expose provider keys and sensitive traffic and can also let attackers tamper with prompts/responses processed by downstream agents. BerriAI’s fix is included starting with LiteLLM v1.83.14-stable—upgrade to that release or later to mitigate.

Researchers at Varonis Threat Labs describe a chained vulnerability in Microsoft 365 Copilot Enterprise Search that could let attackers steal emails, calendar information, and indexed files after a user clicks a seemingly legitimate microsoft.com link. The issue is tracked as CVE-2026-42824 and matters because it can also expose time-sensitive authentication material such as one-time codes and MFA codes, potentially enabling account takeover. Microsoft has issued a critical mitigation on its backend, but tenant admins still need to monitor for suspicious Copilot Search URLs and related outbound requests.

A newly detailed attack chain, dubbed SearchLeak, shows how Microsoft 365 Copilot Enterprise can be turned into a one-click data theft mechanism using a specially crafted URL. The technique targets sensitive content from a victim’s mailbox, OneDrive, or SharePoint by abusing Microsoft 365 Copilot Search, ultimately exfiltrating data via browser-driven and Bing-facilitated requests. Microsoft addressed the issue with fixes for CVE-2026-42824, rated critical, which is important because exploiting the weaknesses does not require additional user actions beyond clicking the link.

Attackers are actively exploiting a critical remote code execution flaw in Jenkins, tracked as CVE-2026-53435. The issue allows adversaries to run arbitrary code on Jenkins controllers, including impersonating users and reaching the Script Console to execute commands or access sensitive files. Jenkins users on Jenkins 2.567 and earlier, as well as LTS 2.555.2 and earlier, should patch immediately; the same advisory also addresses open-redirect issues CVE-2026-53436 and CVE-2026-53437, though they are less severe.