< ciso
brief />
Nobitex Sanctions, Frontier AI Review Order and Redis RCE Flaw

Nobitex Sanctions, Frontier AI Review Order and Redis RCE Flaw

Coverage: 03 Jun 2026 (UTC)

< view all daily briefs >

Policy actions shaped the security agenda on 3 June 2026, as the U.S. Treasury sanctioned Iran's largest cryptocurrency exchange and a new executive order set up a voluntary pre-release security review for frontier AI models. Vulnerability disclosures spanned Redis, HTTP/2 server implementations, Acer routers and Microsoft developer and mobile software. Researchers also documented a supply-chain compromise of Red Hat npm packages, new botnet and remote access trojan campaigns, and active targeting of fuel tank monitoring systems, while Microsoft and AWS shipped tooling to contain, train and run AI agents.

Sanctions and Frontier AI Policy

The U.S. Treasury's Office of Foreign Assets Control designated Nobitex, Iran's largest cryptocurrency exchange, for facilitating transactions tied to terrorist activities, sanctions evasion and payments linked to the Islamic Revolutionary Guard Corps (IRGC). The action names several senior executives and founders and forms part of the "Economic Fury" campaign, which simultaneously targeted three other Iranian exchanges: Wallex, Bitpin and Ramzinex. OFAC cited Chainalysis data indicating that Iran's crypto ecosystem received about $7.8 billion in 2025, with IRGC-associated addresses accounting for more than half of the value received in Q4 2025. Nobitex reportedly processed over 50% of Iranian digital asset inflows in 2025, compared with roughly 12% for Wallex and 10% for Bitpin. The designations freeze assets under U.S. jurisdiction and bar U.S. persons from transacting with the parties, with the stated aim of disrupting channels used by Iranian regime actors and associated ransomware groups to move value across borders. The exchange was previously hit by a June 2025 breach in which a pro-Israel hacking group claimed to have stolen around $90 million.

Separately, President Trump signed an executive order on June 2 creating a voluntary process under which developers of the most powerful AI models can give US agencies access to their systems for cybersecurity review up to 30 days before broader release. The NSA, CISA and NIST are instructed to develop a classified benchmark defining a "covered frontier model," while the order explicitly prohibits any mandatory licensing or preclearance regime. The measure responds to concerns that advanced models can discover and exploit software vulnerabilities at scale, a risk highlighted by the capabilities of Anthropic's Claude Mythos Preview. The order also directs agencies to strengthen national security, military and civilian federal systems within 30 days, charges CISA with issuing binding directives to expand AI-enabled defensive capabilities and broaden access for smaller operators such as rural hospitals and utilities, and places the Treasury in charge of a new AI cybersecurity clearinghouse for coordinated vulnerability scanning, validation and patching. Industry reaction was cautiously supportive, with some leaders stressing that voluntary schemes need clear intake, timelines and safe-harbor terms to create accountability.

Securing and Scaling AI Agents in the Cloud

At its Build conference, Microsoft unveiled a set of agent security tools intended to reduce the risks of autonomous AI agents in development workflows. Microsoft Execution Container (MXC) is a policy-driven sandbox runtime that confines agentic workloads and enforces access controls over files, networks, credentials and other resources across Windows, Linux and macOS, with backends ranging from OS-native sandboxes to full VMs. The Security Multi-model Agentic Scanning Harness (MDASH), which uses over 100 specialized agents to find vulnerabilities, assess exploitability and reduce false positives, is moving from research into broader preview and Defender integration. Agent 365 SDK and Windows 365 for Agents provide development and managed workspace environments integrated with Defender, Entra, Intune and Purview. Two open-source initiatives address governance: ASSERT for spec-driven evaluation and regression testing of agent behavior, and Agent Control Specifications (ACS) for portable governance policies that travel with agents across platforms.

Amazon Web Services made GPT-5.4 on GovCloud generally available through Amazon Bedrock in AWS GovCloud (US-West), targeting government and regulated-industry customers. The model offers native computer-use capabilities and deep reasoning for coding, document understanding and multi-step agentic workflows, and AWS states that customer data remains in-partition and is not used to further train the model. Amazon SageMaker AI also launched multi-turn reinforcement learning, a serverless capability that trains models against a user-provided agent environment and rewards the full sequence of decisions an agent makes, so smaller models can be specialized to match or exceed larger general-purpose models on specific tasks. SageMaker manages rollout orchestration, trajectory collection, training and checkpoints, includes MLflow tracking, and produces reward, pass@k and trajectory metrics; initial model support includes Qwen 3.6 27B, Nova Lite 2.0, GPT-OSS-20B and Gemma 31B in us-west-2, with select models in us-east-1.

On resilience, the Application Recovery Controller gained three Region switch blocks: Amazon Aurora serverless scaling, Amazon Aurora provisioned scaling and Amazon Neptune global database failover. The Aurora blocks automate scaling of scaled-down secondary clusters during failover, removing a manual step that previously added critical minutes to recovery, while the Neptune block consolidates planned switchover and unplanned failover into a single automated plan. All three support cross-account orchestration. AWS Compute Optimizer now offers a 32-day lookback for rightsizing recommendations, up from 14 days, covering EC2 instances, Auto Scaling groups, RDS databases, EBS volumes and ECS services. The capability carries no additional cost and is available in all Compute Optimizer Regions except AWS GovCloud (US) and China.

Amazon ECS Managed Instances now supports Trainium and Inferentia accelerators, including Inferentia2, Trainium1 and Trainium2. Operators create a capacity provider, select accelerated instance types and specify NEURON_CORE=all in the task definition, after which ECS places a single task per instance to dedicate accelerator resources. Usage incurs standard EC2 charges plus a management fee.

Patch Priorities: Redis, HTTP/2 and Edge Devices

Team Xint Code and Wiz detailed CVE-2026-23479, a two-year-old authenticated use-after-free in Redis that allows remote code execution and was found with help from an autonomous AI security tool during a hacking competition. Introduced in Redis 7.2.0, the bug resides in unblockClientOnKey(), where a client structure may be freed and then read. The exploit chain leaks a heap pointer through a Lua EVAL, grooms memory to replace the freed client with a fake structure, and overwrites a Global Offset Table entry so that strcasecmp() points to system(). Many Redis instances run without passwords and default users hold the privileges the chain requires, and the official Docker image's partial RELRO simplifies the final step. Redis reports no evidence of active exploitation, notes that managed Redis Cloud is patched, and has released fixes in versions 7.2.14, 7.4.9, 8.2.6, 8.4.3 and 8.6.3. Interim mitigations include restricting internet exposure, enforcing TLS, tightening ACLs, denying @scripting where unused and rotating shared credentials.

Acer acknowledged two maximum-severity Wave 7 zero-days affecting mesh routers on firmware T7c_GBL_1.01.000055 or earlier, disclosed by researcher Gergo Pap. CVE-2026-49200 exposes an unauthenticated log file containing plaintext web and Telnet credentials, while CVE-2026-49201 stems from a hardcoded AES key in upload.cgi that allows backups to be decrypted, modified and re-encrypted for persistent backdoors. No fixes were available at publication; Acer expects firmware updates by the end of June 2026 and advises disabling remote management or restricting it to trusted IPs in the meantime.

Researchers at Calif disclosed HTTP/2 Bomb, a remote denial-of-service technique affecting NGINX, Apache HTTPD, Microsoft IIS, Envoy and Cloudflare Pingora. The attack combines HPACK compression abuse, where one byte on the wire forces full header bookkeeping on the server, with a Slowloris-like zero-byte flow-control window that prevents those allocations from being freed, allowing a single client to exhaust gigabytes of memory in seconds without triggering decoded-size limits. NGINX 1.29.8 and later add a max_headers directive, and Apache mod_http2 v2.0.41 addresses the issue; disabling HTTP/2 is the alternative for both. IIS, Envoy and Pingora had no patches at the time of reporting.

Enclave Security reported FlagLeft, a development flag left enabled in a shared Microsoft SDK that caused Word, PowerPoint, Excel, Microsoft 365 Copilot, Loop and OneNote on Android to skip verification restricting FOCI family refresh tokens to trusted Microsoft apps. Any app on the same device could obtain a user's token to read email, open files and send messages; Teams was not affected. Microsoft issued four CVEs on May 12 and shipped patched builds via Google Play, with no public evidence of exploitation. Users are advised to update and consider revoking and rotating refresh tokens on devices that ran vulnerable builds. In a separate disclosure, researcher Ammar Askar released proof-of-concept code for an unpatched VS Code zero-day that abuses webview message passing to install an extension and capture the GitHub OAuth token posted by github.dev, which can enumerate all private repositories a victim can access. The issue has no CVE yet; clearing cookies and local site data for github.dev restores the sign-in prompt that blocks silent token reuse.

Supply Chain, Botnets and Critical Infrastructure Targeting

Microsoft Threat Intelligence documented the Miasma npm campaign, in which attackers tampered with the RedHatInsights/javascript-clients CI/CD pipeline and published 32 malicious packages across more than 90 versions in the @redhat-cloud-services namespace. The packages carried legitimate provenance signatures via GitHub Actions OIDC and used an npm preinstall hook to run an obfuscated 4.29 MB dropper that downloads the Bun runtime and a secondary payload. The malware harvests GitHub Actions runner memory, SSH keys, AWS, Azure and GCP CLI credentials, HashiCorp Vault data, Kubernetes secrets, browser and wallet data, and Anthropic API keys, then propagates by committing code to unprotected branches or republishing packages with forged SLSA provenance. Microsoft coordinated removals with npm, published Defender detections and KQL hunting queries, and recommends rotating credentials and hardening CI/CD publishing controls.

FortiGuard Labs identified C0XMO, a Gafgyt botnet variant that exploits a DD-WRT UPnP stack buffer overflow (CVE-2021-27137) to deliver multi-architecture payloads. The variant splits scanning into a separate Python script that brute-forces Telnet and SSH, uses HTTP- and ADB-based exploits, and avoids honeypots via blacklists. The bot removes rival botnets, persists through cron and shell profiles, and supports 19 DDoS methods. Recommended defenses include updating firmware, disabling Telnet and UPnP where unnecessary, enforcing strong credentials and monitoring for outbound scans.

An analysis of the Canvas cyberattack revisited the early May 2026 compromise of Instructure's Canvas LMS, attributed to ShinyHunters. The attackers defaced login pages and claimed to have exfiltrated 3.65 terabytes of data tied to roughly 275 million individuals at nearly 9,000 educational institutions. Instructure said a support-ticket vulnerability in its Free for Teacher environment was exploited and disabled that service. The piece highlights concentrated SaaS risk, the danger of lower-security auxiliary environments outside compliance scopes, and the need for vendor risk assessments to cover incident response maturity, data segmentation and communications; congressional inquiries have requested briefings from Instructure leadership.

Huntress described a malspam campaign that uses DoubleClick redirects as an initial redirector to deliver DesckVB RAT, a .NET remote access trojan. HTML attachments lead through personalized landing pages to a ZIP archive whose JavaScript and PowerShell stages fetch a loader that disables security controls, injects the RAT into Microsoft-signed processes, adds Defender exclusions and patches AMSI and ETW. Suggested mitigations include forcing script files to open in Notepad via GPO, enforcing DMARC, DKIM and SPF, and sandboxing attachments and links at the email gateway.

CISA, with the FBI, NSA, the Department of Energy and partners, warned that internet-exposed automatic tank gauge (ATG) systems used across Energy, Chemical, Food and Agriculture, and Transportation sectors are being actively targeted. Observed vectors include authentication bypasses, hardcoded credentials, command execution, SQL injection and privilege escalation, enabling attackers to alter tank volumes, tamper with pump controls and disable alerts. The agencies have not attributed the activity, though reporting has linked similar intrusions to Iranian-affiliated actors. Operators are urged to remove ATGs from direct internet access, restrict remote connectivity, replace default passwords, enable multifactor authentication, apply updates and monitor for unauthorized changes. Proofpoint, meanwhile, reported that the Chinese-speaking, financially motivated group TA4922 has expanded into Europe and South Africa using new Atlas RAT malware, delivered via payroll, tax and HR lures as well as WhatsApp, LINE and Microsoft Teams contact. The toolset also includes RomulusLoader, SilentRunLoader and Winos4.0/ValleyRAT, and code artifacts suggest possible use of large language models in development; Proofpoint published indicators of compromise and C2 infrastructure.

Nobitex Sanctions, Frontier AI Review Order and Redis RCE Flaw · CISO Brief