An empirical study evaluated the network exposure of Ray control-plane components in distributed large language model (LLM) inference deployments on Kubernetes using Amazon EKS 1.35. The study found that multiple Ray control-plane sockets were not declared in pod metadata, allowing an unprivileged pod in a different namespace to access Ray GCS and raylet RPC services over unauthenticated cleartext gRPC. Default security scanners did not detect this exposure. A default-deny NetworkPolicy blocked access in steady state, and WireGuard encryption added modest overhead. The Ray Job API responded to unauthenticated requests on CPU Ray images but was not active in GPU deployments. The findings reflect vendor default configurations rather than new vulnerabilities, and a related CVE affecting vLLM was already fixed upstream.
An empirical evaluation of NVIDIA OpenShell v0.1.2 sandbox policy enforcement was conducted on macOS using a local microVM driver. The study tested 123 trials against documented egress controls and found no bypass of documented controls. However, egress was observed through operator-opened paths such as read-write rules, query and header values on GET-only rules, audit-mode rules, and auto-approval mechanisms. The evaluation highlights limitations in policy enforcement but no security bypasses were confirmed. The test environment and logs are publicly available for verification.
A technical audit evaluated the default security configurations of 15 official Helm charts used for AI serving, vector databases, and Model Context Protocol (MCP) agents on Kubernetes. Key findings include plaintext secret handling, unauthenticated remote code execution, and over-privileged agent access due to insecure default settings. Standard static analysis tools failed to detect some issues in Custom Resource Definitions. The report provides reproducible test commands, network logs, and remediation advice. No active exploits are reported, and the findings are based on local test cluster probes.
CVE-2026-79417 is a local denial-of-service vulnerability in Argus Monitor caused by an exposed IOCTL that allows unprivileged users to disable the x86 MONITOR and MWAIT instructions. Exploiting this vulnerability requires leveraging a time-of-check to time-of-use (TOCTOU) bug related to the SeLocateProcessImageName function. Successful exploitation triggers a HYPERVISOR_ERROR bugcheck, impacting Hyper-V and other kernel components that rely on these instructions.
RCEKit is an open-source toolkit designed to detect and confirm remote code execution (RCE) vulnerabilities during authorized penetration testing and security research. It uses multiple methods to distinguish confirmed RCE from weaker signals or false positives by generating unique tokens and verifying their presence in target responses or out-of-band callbacks. The tool has been tested against real-world CVEs such as Webmin CVE-2019-15107, Apache Struts2 S2-001, and Log4Shell CVE-2021-44228, demonstrating its ability to confirm or classify RCE findings accurately. RCEKit does not report uncertain cases as confirmed, instead assigning them lower confidence tiers. It supports testing via URLs or captured HTTP requests and offers various probing methods, including reflected, eval, time-based, lookup, and deserialization injection techniques. The tool is intended for use only on authorized targets and requires careful configuration due to the high volume of requests it generates.
A file collision bug in the PcapSplitter library causes silent packet loss when TCP sessions reuse the same 5-tuple, resulting in file truncation or corruption. This occurs because the filename generation is based only on IP and port, causing multiple sessions to overwrite the same output file. The issue was identified during PCAP processing and fixed by suffixing filenames only on actual collisions. The bug led to fewer packets being written than reported, with no error exit codes, making detection difficult.
A security researcher named Gal Weizman disclosed a browser security research achievement involving a single browser extension that successfully exploited vulnerabilities in Chrome, Comet, Edge, Opera, and Claude in Chrome. This research led to the discovery of two CVEs and earned $20,000 in bounty rewards from multiple major vendors including Anthropic, Perplexity, Google, Microsoft, and Opera. No detailed technical information or affected versions are provided.
This content is a Reddit post requesting participation in a survey for thesis work, containing a link to a Qualtrics survey form. There is no indication of a security threat, vulnerability, or malicious activity associated with this request.
A Chinese-speaking threat actor group known as Red Heron exploited a recently disclosed remote code execution vulnerability (CVE-2026-60004) in Gitea, a self-hosted Git service, in a multinational campaign. The campaign targeted internet-facing Gitea instances across multiple countries, including Canada, Argentina, Taiwan, the United States, and Sri Lanka, focusing on sectors such as defense, elections, energy, aerospace, telecommunications, government, and research. The attackers used automated tools to steal source code, credentials, and maintain persistent access, including root-level control on some infrastructure. They deployed a novel Linux implant named JITTERLY with extensive post-exploitation capabilities and embedded a previously undocumented rootkit called SIXZUT to maintain stealth and persistence. The campaign demonstrates rapid weaponization of n-day vulnerabilities in development platforms and highlights significant risks to source code confidentiality and infrastructure integrity.
EchelonGraphBot is a web crawler operated by EchelonGraph, Inc. that scans public websites to check configurations such as HTTPS enforcement, certificate expiration, TLS versions, security headers, cookie flags, and redirects. It visits a large number of third-party hosts approximately once per day, respecting robots.txt and rate limits. Operators can opt out via a DNS TXT record or robots.txt. The bot's activity is transparent and publicly documented, with no evidence of malicious intent or exploitation. The operator offers to adjust probing frequency upon request and honors opt-out requests.
This report discusses a method to locate Flutter's TLS certificate verifier function within a stripped libflutter.so binary on ARM64 devices without relying on fragile byte signatures. Flutter bundles its own BoringSSL and does not use Android's network security config, requiring patching of the ssl_crypto_x509_session_verify_cert_chain function to modify trusted CAs. The function can be identified by cross-referencing unique strings and specific ARM64 prologue characteristics, enabling reliable patching across multiple Flutter engine versions. This technique was validated on apps using Flutter engine versions 2.19.2 through 3.11.1. The report also highlights a common pitfall in patching tools that fail to handle split APKs correctly. No active exploits or vendor patches are mentioned.
CVE-2026-84388 is a critical vulnerability in the FortiPAM Chrome extension used for Privileged Access Management. It allows any website visited by the user to manipulate the browser's proxy settings for the session and to open new tabs that can be screen recorded and streamed to an attacker's server. This enables attackers to conduct phishing attacks by capturing sensitive user activity in attacker-controlled tabs without user consent.
A Redis cryptomining botnet compromised 3,562 Redis servers by exploiting unsecured no-auth configurations. The botnet operator's own files were exposed in an open directory, revealing the full toolkit and detailed campaign logs. The attack leveraged rogue replication commands to deploy a cron job that runs the XMRig miner, targeting Monero mining pools. The issue is due to missing authentication and insecure default configurations, not a software vulnerability. The affected Redis versions range from 2.8.17 to 7.2.0. Mitigation involves configuring Redis securely by enabling authentication and disabling replication features if unused.
A large-scale phishing campaign used invisible Unicode characters from the deprecated Unicode Tags block to obfuscate keywords related to loans and financing, bypassing email filters that rely on contiguous ASCII string matching. The campaign sent between 1 million and 2.37 million messages daily at its peak in early 2026, primarily targeting recipients with fraudulent loan offers. The messages were routed through the legitimate marketing platform ActiveCampaign, leveraging its reputation to evade detection. There is no software vulnerability or patch associated with this technique; it exploits differences in how email clients and filters process Unicode text. Mitigation requires enhanced email filtering that inspects raw Unicode content and flags or removes invisible characters to detect obfuscated phishing keywords.
On August 31, 2026, a malicious Terraform module was served from the legitimate registry.coder.com domain for approximately 14 hours. This module included an extra telemetry block that exfiltrated environment data to an attacker-controlled domain. No CVE was assigned, so standard vulnerability scanning tools did not flag this incident. The compromised module affected users who created or updated templates, ran template dry-runs, or built workspaces with module caching disabled during the incident window. Sensitive data potentially leaked includes cloud credentials, API keys, tokens, SSH keys, and deployment configurations. Coder has released patched builds and a tool to help identify affected deployments, but detection requires manual checking. No attacker-side logs are available, so affected users must proactively verify exposure.
An authentication bypass vulnerability affects Proxmox VE 7 releases that are end-of-life (EOL). The issue allows unauthorized access if port 8006 is exposed, potentially leading to system compromise. The vulnerability has been publicly discussed on the Proxmox forum and Reddit but lacks official vendor advisory or patch information. Exploitation demonstrations have been shared, but no confirmed active exploits in the wild are reported. The affected versions are implied to be Proxmox VE 7.0 through versions prior to 8.0.4, but exact affected versions are not explicitly confirmed.
PATCHCORD is a newly identified malware cluster targeting Afghan telecom providers and critical infrastructure organizations in South Asia. The malware is a custom backdoor implant written in C/C++ and delivered via sector-specific lures such as fake VPN installers and telecom management tools impersonating legitimate Afghan Telecom software. This campaign is ongoing and was discovered by the Acronis Threat Research Unit. The malware aims to infiltrate sensitive telecom and infrastructure environments in the region.
A self-hosted EDR agent named Sentora had multiple critical authentication and authorization bugs, including an unauthenticated remote uninstall endpoint accessible to anyone on the LAN. The root cause was a common design flaw where the system trusted caller-supplied identity from various sources such as URL paths, headers, and request bodies without proper validation. Other issues included permissive authentication accepting any non-empty key, unauthenticated automation reporting that could fake containment success, LDAP injection in login, and missing authorization checks on multiple routes. The author discovered these issues through code review and automated tests, and has since fixed them on the main branch.
AWSHound is a free, self-hosted tool that generates attack path graphs for AWS accounts and organizations using BloodHound CE. It analyzes IAM policies, boundaries, SCPs, RCPs, and resource policies offline to visualize potential compromise paths. This is a security tool release rather than a vulnerability or active threat.
A security researcher reported escaping the WebAssembly sandbox, achieving arbitrary shell execution on the host system. The issue relates to WABT's wasm2c tool, which compiles WebAssembly to C code assumed to maintain sandbox security. This proof-of-concept demonstrates that the sandboxing guarantees can be bypassed, potentially allowing code execution outside the intended WebAssembly environment.
