< ciso
brief />
Exploited Fortinet, Cisco and cPanel Flaws; AI Guardrails Expand

Exploited Fortinet, Cisco and cPanel Flaws; AI Guardrails Expand

Coverage: 16 Jun 2026 (UTC)

< view all daily briefs >

Active exploitation shaped the vulnerability picture on June 16, with CISA adding a LiteSpeed cPanel plugin flaw to its Known Exploited Vulnerabilities catalog and attacks reported against Fortinet FortiSandbox and Cisco Catalyst SD-WAN Manager. Threat research detailed DragonForce ransomware operators hiding command-and-control traffic in Microsoft Teams relays and a Windows expansion of the China-linked SprySOCKS backdoor. Vendors and industry groups focused on runtime controls for agentic AI and coordinated patching of open-source software. In policy, the UK outlined mandatory age verification for new social media accounts.

Runtime Controls for Agentic AI and Open-Source Defense

Chainguard announced the Athena coalition on June 16, bringing together dozens of partners, including BNY, Cisco, Cloudflare, Docker, JPMorganChase and PwC, to defend open-source software against vulnerabilities discovered by frontier AI models. The platform aggregates vulnerability intelligence from members, uses frontier AI tools to analyze findings, and lets Chainguard privately patch and rebuild affected projects as hardened versions distributed to members through Chainguard Libraries. Where patches are not yet ready, members operating infrastructure and security layers deploy non-patch mitigations, while cybersecurity partners contribute detections, signatures and virtual patches. Disclosure is coordinated upstream so fixes become part of the public project tree. Athena has processed more than 20,000 findings and shipped over 2,000 patches across some 500 open-source projects, with the first wave of public disclosures scheduled for July. Chainguard plans to collaborate with the Linux Foundation on a Security Incident Response Team and a maintainer-of-last-resort program, and frames Athena as an AI cybersecurity clearinghouse, citing recent US government AI security directives and the decision to restrict access to Anthropic’s Mythos.

AWS released the InvokeGuardrailChecks API for Amazon Bedrock Guardrails, aimed at agentic applications that take many iterative steps with differing risk profiles, where a single global guardrail is insufficient. The API is resourceless and runs in detect-only mode: callers specify which safeguards to run at each step of an agent loop and receive numeric severity and confidence scores to support custom thresholds and automated actions such as blocking, retrying, passing or logging. Independent safeguards cover content filtering (hate, violence, sexual content, insults, misconduct), prompt attack detection with jailbreaks, prompt injection and prompt leakage exposed as separate checks, and detection of supported PII types. No guardrail IDs or version management are required. The capability is available in US East (N. Virginia), US East (Ohio), US West (Oregon), Europe (London), Europe (Stockholm), Asia Pacific (Tokyo) and Asia Pacific (Sydney).

In a related update, AWS extended Automated Reasoning checks in Bedrock Guardrails to the Asia Pacific (Sydney) Region, adding to prior availability in multiple US and European regions. The feature uses formal verification techniques rather than sampling-based testing to validate LLM outputs, and AWS reports up to 99% accuracy in detecting correct responses. It is aimed at regulated industries such as finance, healthcare and legal, with use cases including pre-production validation, enforcement of business rules and quality assurance to reduce risks from hallucinations and policy noncompliance. It is accessible through the Amazon Bedrock console and SDK.

Palo Alto Networks and Databricks announced an integrated runtime security offering that connects Prisma AIRS with the Databricks Unity AI Gateway. The gateway acts as a central control plane that routes AI activity to Prisma AIRS for real-time inspection of prompts, responses and tool calls across models, agents, applications and MCP servers. The integration targets prompt injection, exfiltration of sensitive data (PII, PHI, PCI), malicious code and abusive URLs, and can redact or block data according to corporate policy. The partners argue for a shift from reactive controls to a fail-closed runtime posture and describe a Security-as-Code experience for developers, while noting in the announcement that outcomes are subject to risks and uncertainties.

Exploited Flaws in Hosting, Sandbox and SD-WAN Platforms

CISA added a high-severity flaw in the LiteSpeed cPanel user-end plugin to its KEV catalog and ordered Federal Civilian Executive Branch agencies to remediate within three days under Binding Operational Directive 26-04. Tracked as CVE-2026-48172 (also referenced with CVE-2026-54420) and reported by Namecheap, the bug is a UNIX symlink-following weakness in plugin versions prior to 2.4.8. An attacker with FTP or web shell access can escalate privileges to root on shared hosting environments running CloudLinux/CageFS, creating a critical risk for hosted tenants. LiteSpeed flagged the issue as actively exploited in early June, issued urgent updates and supplied a forensic grep command administrators can use to check for signs of compromise. BOD 26-04 directs agencies to prioritize KEV-listed flaws, publicly exposed assets and vulnerabilities that enable automated or high-impact takeover; CISA also advises following cloud-specific guidance or discontinuing affected services if no mitigations are available.

Defused Cyber observed exploitation attempts within a 24-hour window against three FortiSandbox flaws: CVE-2026-39813, a path traversal in the FortiSandbox JRPC API; CVE-2026-39808, an operating system command injection; and CVE-2026-25089, an OS command injection affecting the FortiSandbox, FortiSandbox Cloud and FortiSandbox PaaS web UI. All three carry a CVSS score of 9.1 and could allow unauthenticated attackers to bypass authentication or execute unauthorized commands via crafted HTTP requests. Fortinet patched the first two in April 2026 and released a fix for CVE-2026-25089 the previous week. Defused noted that the emerging exploit for CVE-2026-25089 appears to have been developed with assistance from an AI model and is currently faulty, and a reliable public exploit has not been disclosed. The activity follows exploitation of FortiClient EMS (CVE-2026-35616), which received out-of-band patches in April 2026. FortiSandbox users are advised to apply vendor updates, verify mitigations and monitor for suspicious activity.

Cisco released fixes for CVE-2026-20262, a medium-severity flaw in the web UI of Cisco Catalyst SD-WAN Manager (formerly SD-WAN vManage) discovered during internal testing and observed in limited active exploitation in June 2026. Inadequate validation of user-supplied input during a file upload process lets an authenticated remote user with write privileges create or overwrite arbitrary files by sending crafted HTTP requests to an API endpoint, potentially enabling privilege escalation to root. On-premises, cloud, Cisco-managed and FedRAMP deployments are affected, and fixed releases include 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. Cisco published indicators of compromise and recommends auditing /var/log/nms/vmanage-server.log for suspicious WAR uploads, noting that follow-on activity may not always produce consistent log artifacts. This is the eighth SD-WAN-related Cisco flaw flagged as actively exploited in 2026, some tied to the APT actor UAT-8616. CISA added the vulnerability to its KEV catalog and required federal agencies to apply mitigations by June 29, 2026.

Unit 42 disclosed a critical vulnerability in the Google Cloud Vertex AI SDK (google-cloud-aiplatform) affecting versions 1.139.0 and 1.140.0. When users did not supply a custom staging bucket, the SDK constructed a predictable bucket name from project and region and used it without verifying ownership. An attacker who knows a victim’s project ID could pre-create that bucket in their own project, grant broad IAM permissions, and use a triggered Cloud Function to replace uploaded model artifacts within a race window of about 2.5 seconds, before the Vertex AI service agent reads them. A malicious model serialized with joblib/pickle abusing the __reduce__ mechanism would then execute arbitrary Python code on deserialization, enabling remote code execution, data exfiltration and lateral movement. Google released a fix in v1.148.0 on April 15, 2026. Unit 42 recommends upgrading and notes that LLM-assisted code review accelerated discovery.

Ransomware and Espionage Operators Refine Stealth

Researchers from Symantec and Carbon Black detailed a 2025 intrusion in which DragonForce ransomware operators remained inside a major US services firm for up to two months. A Go-based remote access trojan dubbed Backdoor.Turn obtained anonymous Microsoft Teams visitor tokens and used legitimate TURN relay servers, encapsulating QUIC sessions so defenders observed only outbound connections to legitimate Teams endpoints, which thwarted detection by standard security products. The intrusion likely began through exploitation of an SQL/MSSQL server vulnerability and also leveraged an undocumented Huawei driver vulnerability later disclosed by Huntress. To persist and move laterally, the attackers removed the Limit Blank Password control, created user accounts and modified firewall rules, combining remote code execution, network scanning, credential theft and credential-based lateral movement before exfiltrating data and deploying DragonForce ransomware.

ESET identified two previously undocumented Windows variants of the SprySOCKS backdoor, until now considered a Linux-only toolset of a China-linked espionage cluster. The WIN_DRV variant uses kernel drivers, RawWNPF and an encrypted DriverLoader, to conceal network connections, processes, files and registry keys, and implements TCP traffic diversion to obscure its true listening port. WIN_PLUS abuses the Windows Print Spooler to run a first-stage loader as a print processor and injects a loader into svchost.exe. Both retain the original C&C protocol, encryption and command set, support TCP, UDP and WebSocket channels, and offer more than 30 operator commands for system enumeration, service management, file operations, SOCKS proxy initialization and interactive consoles. ESET links the activity to the FishMonger cluster, associated with the Winnti umbrella and previously attributed by others to Earth Lusca/Aquatic Panda. Artifacts suggest deployments against government organizations in Honduras, Taiwan, Thailand and Pakistan between 2023 and 2024, and limited indications point to a UEFI bootkit exploiting CVE-2023-24932 in some intrusions.

Console Access, Cloud Platform Updates and UK Age Checks

AWS added support for resource-based policies and resource control policies for AWS Management Console sign-in. Resource-based policies apply at the individual account level, while RCPs operate organization-wide through AWS Organizations; both are evaluated at sign-in and whenever a console session requests new credentials, ensuring continuous enforcement of network-based access restrictions. Combined with AWS Management Console Private Access, administrators can control both the networks users may sign in from and the accounts they may access. The feature is available at no additional cost in all AWS commercial Regions.

Amazon S3 introduced object annotations, a metadata capability that attaches business context to objects in JSON, XML or YAML at up to 1GB per object. Annotations inherit S3 durability and consistency guarantees, move with objects during copy and replication, and are removed when the object is deleted. They can be surfaced in S3 Metadata as read-only Apache Iceberg tables queryable by Amazon Athena, and searched in natural language through agents in Amazon SageMaker Unified Studio or the S3 Tables MCP server. Separately, S3 Vectors pricing now lowers data-processed charges by up to 80% for queries on indexes larger than 10 million vectors, applied automatically in all Regions where the service is available.

AWS Transform for mainframe gained a traceable reimagine workflow spanning portfolio assessment, requirements generation and code production for z/OS COBOL and PL/I workloads, with every requirement and generated artifact linked back to the original source code and MCP-based connectors for IDEs including Kiro. AWS also made EC2 P6-B200 instances available in Asia Pacific (Mumbai); each p6-b200.48xlarge includes eight NVIDIA Blackwell GPUs with 1440 GB of GPU memory and up to 3.2 terabits per second of EFAv4 networking, running on the AWS Nitro System.

The UK government announced on June 15 plans to ban under-16s from social media and require age verification for new accounts, with regulations scheduled before Christmas and implementation in spring 2027. The rules target platforms offering social interaction and algorithmic feeds, naming Instagram, TikTok, YouTube, Snapchat, Facebook and X, while excluding messaging apps such as WhatsApp and Signal; long-standing accounts are generally grandfathered. New users will likely need to upload ID or use a facial age scan. Security and privacy researchers warn that the mechanisms are technically weak, easily circumvented via VPNs, and create data-breach and identity-theft risks by centralising ID and biometric data. Ofcom has been asked to run a rapid study on verification options, while critics including the Open Rights Group argue the approach normalises intrusive verification and undermines online anonymity.