< ciso
brief />
Gemini CLI RCE Fix, Linux Copy Fail and Agentic AI Security Updates

Gemini CLI RCE Fix, Linux Copy Fail and Agentic AI Security Updates

Coverage: 30 Apr 2026 (UTC)

< view all daily briefs >

Three high-priority patch items lead 30 April 2026: a maximum-severity remote code execution flaw in Google's Gemini CLI for headless CI pipelines, a portable Linux kernel privilege escalation known as Copy Fail, and a cPanel and WHM authentication bypass exploited as a zero-day since February. Threat research documented a credential-stealing compromise of PyTorch Lightning on PyPI, two SaaS-focused extortion groups and malicious GenAI browser extensions. Hyperscalers and security vendors also shipped agentic AI controls alongside new hardware trust, post-quantum and key management capabilities.

Critical Fixes for CI Tooling, the Linux Kernel and cPanel

Google has issued Gemini CLI fixes for a remote code execution vulnerability rated CVSS 10.0 in @google/gemini-cli and the google-github-actions/run-gemini-cli workflow, following research by Novee Security. The issue, which was not assigned a CVE in the advisory, affected @google/gemini-cli releases prior to 0.39.1 and certain preview builds, as well as run-gemini-cli versions prior to 0.1.22. The risk is specific to headless CI usage, where Gemini previously auto-trusted workspace folders. An attacker able to plant crafted configuration files or environment variables, for example under a local .gemini/ directory, could have them loaded without review or sandboxing and trigger command execution on the host. Google now requires explicit workspace trust before configuration files are read. Maintainers can set GEMINI_TRUST_WORKSPACE to 'true' for workflows that only process trusted inputs and should follow the vendor's hardening guidance where untrusted inputs are possible. Version 0.39.1 also hardens tool allowlisting for the --yolo mode so that automatic approval can no longer bypass configured allowlists; administrators are advised to review those allowlists because some workflows may need adjustment.

The same research covered two Cursor issues. CVE-2026-26268 (CVSS 8.1), affecting Cursor versions prior to 2.5, allows a malicious embedded bare .git repository with a post-checkout hook to execute code when an autonomous Cursor agent performs a checkout. A second flaw rated CVSS 8.2, dubbed CursorJacking, can let installed extensions access sensitive credentials stored in a local SQLite database because of missing access-control boundaries. Recommended steps are to update Cursor and avoid untrusted extensions.

In the Linux kernel, a local privilege escalation tracked as CVE-2026-31431 (CVSS 7.8) and nicknamed Copy Fail stems from a logic error in the algif_aead module of the cryptographic subsystem, introduced by a 2017 in-place optimization. Combined with the AF_ALG socket interface and the splice() system call, the flaw lets an unprivileged process perform a controlled four-byte write into the page cache of any readable file, including a setuid binary such as /usr/bin/su, which can then be executed to run injected code as root. Theori, which found the bug using its AI-driven Xint Code platform, reported it to the kernel security team on March 23, and upstream fixes were merged in early April, with kernel releases such as 6.18.22, 6.19.12 and 7.0 cited. The published proof of concept is a 732-byte Python script that researchers describe as reliable, requiring no timing races or kernel memory offsets, and it reportedly roots distributions including Ubuntu 24.04 LTS, Amazon Linux 2023, RHEL 10.1 and SUSE 16. Although not remotely exploitable on its own, the bug has cross-container implications because the page cache is shared across processes and containers on the same host. Amazon Linux, Debian, RHEL, SUSE and Ubuntu have published advisories. Where patching is delayed, Theori recommends blocking AF_ALG socket creation or disabling the algif_aead module, with priority for multi-tenant systems, Kubernetes and container clusters, CI/build runners and cloud services that execute untrusted code.

Hosting operators face CVE-2026-41940, a critical authentication bypass in cPanel, WHM and WP Squared that has been exploited as a cPanel zero-day since late February 2026. Researchers at watchTowr identified a CRLF injection in the login and session-loading processes, where attacker-controlled data from the Authorization header is written into server-side session files before authentication and without proper sanitization. KnownHost reported execution attempts as early as 23 February. cPanel released patches on April 28 for multiple 11.x branches and WP Squared and recommends restarting the cpsrvd service after updating. Rapid7 Shodan scans identify roughly 1.5 million internet-exposed cPanel instances, although the number vulnerable to this specific issue is unknown. Successful exploitation can give an attacker control over the host, its configurations and databases, and any websites it manages. Interim measures include blocking management ports 2083, 2087, 2095 and 2096 and temporarily stopping cpsrvd and cpdavd; Namecheap blocked ports until patches were available. Where indicators of compromise are found, operators should purge sessions, reset all credentials, audit logs and investigate persistence and lateral movement.

Supply Chain, SaaS Identity and Browser Extension Threats

Two malicious releases of PyTorch Lightning, versions 2.6.2 and 2.6.3, were published to PyPI on April 30 in a PyPI supply chain attack aimed at credential theft. The packages contain a concealed _runtime directory with a Python starter script that downloads the Bun JavaScript runtime and runs an obfuscated payload of roughly 11MB named router_runtime.js to harvest tokens and secrets. Stolen GitHub tokens are validated and then used to create or silently overwrite files across repositories, pushing a worm-like payload into as many as 50 branches per writable repository under an identity claiming to be Anthropic's Claude Code. An npm propagation vector injects a postinstall hook into local package.json files, bumps patch versions and repacks archives, which can lead to altered packages being published if a developer publishes from an infected environment. Vendors attribute the activity to TeamPCP and link it to the Mini Shai-Hulud operation, citing overlap with earlier incidents targeting Checkmarx, Bitwarden, Telnyx, LiteLLM and Aqua Security Trivy. PyPI administrators have quarantined the project, and initial indications from maintainers point to a likely compromise of the project's GitHub account. Recommended actions include removing the affected versions, downgrading to 2.6.1, rotating exposed credentials and tokens, auditing npm packages and developer workstations, and monitoring repositories for unexpected commits. Researchers also observed a similar compromise of the npm package intercom-client 7.0.4 consistent with the same campaign.

CrowdStrike Counter Adversary Operations has tracked CORDIAL SPIDER and SNARKY SPIDER since October 2025 in SaaS-centric intrusions focused on large-scale data theft and extortion, as detailed in a CrowdStrike threat analysis. Both groups use voice phishing to steer victims to SSO-themed adversary-in-the-middle pages that capture credentials and live session tokens, then abuse identity provider trust relationships to pivot into connected SaaS applications from a single authenticated session. Persistence is established by enrolling attacker-controlled MFA devices, including Genymobile Android emulators and QEMU-based instances, and in some cases by removing existing MFA. The actors create inbox rules that delete security notifications, search SaaS platforms for terms such as confidential, SSN and contracts, and exfiltrate data from SharePoint, Google Workspace, HubSpot and other services. The intrusions exploit customer configuration weaknesses, such as the absence of phishing-resistant MFA and overly permissive access controls, rather than platform vulnerabilities, and route traffic through commercial VPN and residential proxy providers including Mullvad, Oxylabs and NetNut. CrowdStrike positions Falcon Shield detection and SaaS security posture management as its response.

Browser extensions are the focus of Unit 42 research into 18 Chrome extensions marketed as GenAI productivity tools that operated as malware, delivering remote access trojans, infostealers, search hijackers, brand impersonators and adversary-in-the-browser components. Case studies include an MCP Server extension that maintained a persistent WebSocket connection to a command-and-control server accepting more than 30 remote commands, Supersonic AI, which exfiltrated Gmail and Outlook content and captured verification codes, and Reverse Recruiting, which stole OpenAI, Gemini and Claude API keys. Observed techniques include hooking window.fetch and XHR, passive DOM scraping, dynamic proxy configuration through remote PAC scripts, WebSocket-based command channels and persistence through chrome.storage.sync, cookies and localStorage. Analysts also noted code fingerprints suggesting AI-accelerated malware development. Google removed many of the flagged extensions or warned owners to remediate policy violations. Unit 42 recommends treating browsers as a primary enterprise attack surface, vetting extensions as third-party software, enforcing least-privilege permissions and monitoring runtime network behavior, and has published indicators of compromise including extension IDs and domains.

Agentic AI Platforms Add Guardrails, Identity and Automation

AWS released Neuron Agentic Development, an open-source framework of agents and skills for low-level kernel development on AWS Trainium and AWS Inferentia. The initial release targets the Neuron Kernel Interface (NKI) and enables agent-driven IDEs, with Claude Code and Kiro cited as examples, to author kernels from natural-language descriptions, debug and apply fixes, capture profiles and produce performance analyses that point to hot lines of code. The codebase is available through a public GitHub repository, and AWS notes that agent-produced kernel code still requires human review, testing and validation. Separately, Amazon Quick is previewing Microsoft 365 extensions for Excel, PowerPoint and Word that run AI-driven tasks such as spreadsheet analysis, deck generation from organization-defined templates and contract redlining with track changes enabled. The preview is available in regions including US East (N. Virginia), US West (Oregon), Europe (Ireland), Europe (Frankfurt) and Europe (London), and customers are advised to consider governance, permissions and data residency implications before enabling AI-driven operations inside Microsoft 365.

Microsoft detailed Microsoft Security updates aimed at AI agents, development lifecycles and sensitive data. New Microsoft Defender features in the Agent 365 tooling gateway, in preview, evaluate agent actions through webhooks to detect, block and investigate anomalous or risky behavior before actions are executed. The integration of Defender for Cloud with GitHub Advanced Security is now generally available, linking source code changes to production environments, prioritizing alerts using runtime context and offering AI-assisted remediation. A new Microsoft Purview demo walks through an end-to-end data security investigation using AI-powered deep content analysis and a data risk graph. Microsoft frames the changes within its Zero Trust for AI approach and Secure Future Initiative.

Google Cloud COO Francis deSouza summarized security themes from Next '26 in a Cloud CISO Perspectives post, arguing for multicloud and multi-AI architectures to achieve resilience and avoid single-vendor concentration risk. Google announced roughly 220 product updates and reported that a Gemini-powered Triage and Investigation agent has processed more than 5 million alerts this year, reducing 30-minute manual tasks to 60 seconds, while SOC agents yielded a 90% reduction in threat mitigation time. The Gemini Enterprise Agent Platform introduces Agent Gateway for unified governance, Model Armor for prompt sanitization and Agent Identity for authenticated, least-privilege agent actions, and Google positions the integration of Wiz as extending security across self-hosted and multicloud AI environments. Google also published the whitepaper Building Secure Multi-Agent Systems on Google Cloud and referenced threat intelligence on rising data leak activity in Europe, including the Überfall actor, and the UNC6692 intrusion campaign.

Cloudflare and Stripe launched an integration that lets agents create Cloudflare accounts, start paid subscriptions, register domains and obtain API tokens to deploy applications to production. The co-designed protocol used by Stripe Projects builds on OAuth and OIDC, extended with payment tokenization and account provisioning, and is structured around discovery, authorization and payment steps. Users with existing accounts go through a standard OAuth consent flow, and Stripe applies a default spending limit of $100.00 USD per provider per month for agent-initiated purchases, which users can manage through budget alerts. Cloudflare describes the model as generalizable to any platform with signed-in users acting as the orchestrator.

Palo Alto Networks' Unit 42 Frontier AI Defense announced integration with Anthropic's Claude Security, powered by Opus 4.7, for exposure analysis that maps complex exploit chains, deep-stack code review of large codebases and agentic defense workflows that retain human oversight for validation and risk decisions. The company is also participating in Anthropic's Cyber Verification Program to credential legitimate defensive use of frontier models.

On agent identity, Amazon Bedrock AgentCore Identity now supports On-Behalf-Of token exchange, generally available in 14 AWS Regions. The flow exchanges an incoming user access token for a scoped-down token that combines user and agent identity and targets a specific outbound protected resource, removing separate consent prompts per resource and supporting least-privilege access and attribution. Organizations are advised to review how exchanged tokens are scoped, logged and validated downstream. AWS also previewed AgentCore optimization capabilities that generate recommended system prompts and tool descriptions from production traces, validate them through batch evaluations and run A/B tests that report statistical significance; every recommendation requires explicit customer approval before deployment.

Hardware Trust, Post-Quantum Tunnels and Key Management

Microsoft will open-source Azure Integrated HSM firmware, drivers and software, covering the tamper-resistant hardware security module integrated into each new Azure server. The announcement at the OCP EMEA Summit includes independent validation artifacts, among them an OCP SAFE audit report, and the formation of an OCP workgroup to steer architecture, protocol, firmware and hardware development. The module is engineered to meet FIPS 140-3 Level 3 requirements; keys are generated, stored and used entirely within the HSM so they never appear in host memory, guest memory or software processes. It supports TDISP for secure binding with confidential computing environments and complements Azure Key Vault and Azure Managed HSM. Microsoft says the HSM will be broadly available in Azure V7 virtual machines in the coming weeks.

Cloudflare made post-quantum IPsec generally available for its IPsec WAN service, implementing the hybrid ML-KEM approach in draft-ietf-ipsecme-ikev2-mlkem (FIPS 203) to counter harvest-now-decrypt-later attacks. The handshake combines a classical Diffie-Hellman exchange with an ML-KEM exchange to derive session keys without specialized hardware. Interoperability has been validated with Cisco 8000 Series routers (v26.1.1+) and Fortinet FortiOS 7.6.6+, after earlier testing against strongSwan. Cloudflare notes that RFC 9370 did not prescribe specific ciphersuites, leading to nonstandard early vendor implementations, and calls for convergence on interoperable implementations and post-quantum authentication standards for IPsec. The capability is offered at no extra cost, and the company targets full post-quantum readiness by 2029.

In key management, AWS Payment Cryptography now supports Physical Key Exchange, a PCI PIN and P2PE-compliant method for paper-based key exchange in which trained AWS key custodians conduct ceremonies in AWS-operated secure facilities, removing the need for customers to maintain HSMs and key loading devices for infrequent ceremonies; AWS states that electronic key exchange remains the preferred approach. AWS also added OpenSearch index-level encryption, allowing per-index encryption at rest with separate AWS KMS customer managed keys for domains running OpenSearch 3.3 or later in 14 Regions at no additional cost. The feature supports tenant-specific and regulatory key separation, while requiring attention to key policies, IAM controls, rotation and backup and restore processes.

Gemini CLI RCE Fix, Linux Copy Fail and Agentic AI Security Updates · CISO Brief