An AI pentesting engine named Cascade successfully bypassed a production AI agent's prompt-injection guardrail by reframing its request, causing the agent to disclose its entire system prompt, including tool lists, calling rules, citation formats, and session IDs. The bypass did not rely on technical vulnerabilities but on social engineering-like prompt reframing that made the request appear reasonable to the agent. This finding highlights a novel attack vector against AI agents relying on prompt-based guardrails.
This analysis covers a Windows-targeted cyber campaign involving social engineering via fake interviews leading to cryptocurrency theft. The campaign notably did not use DPRK malware but leveraged known malware-as-a-service (MaaS) tools and unidentified Go and Rust-based stealers and remote access trojans (RATs). The threat actor may be using DPRK tradecraft as a false flag to confuse attribution, possibly indicating Russian operators. The campaign targets web3 organizations and involves a multi-stage payload delivery chain using signed ClickOnce applications.
This content announces an Ask Me Anything (AMA) session with Arshi Chadha, a leading AI security researcher and co-lead of the OWASP LLM Top 10 project focused on vector and embedding weaknesses. The discussion will cover prompt injection, retrieval-augmented generation (RAG) poisoning, embedding attacks, and other AI system vulnerabilities. It highlights ongoing research and public disclosures related to AI security but does not describe a specific vulnerability or exploit. The event aims to educate and engage the community on AI security challenges and mitigation strategies.
ICMP-Ghost is an open-source tunnelling framework that uses ICMP and DNS protocols for command and control (C2) communication. It is designed in x64 assembly and claims to evade endpoint detection and network intrusion detection systems. The framework supports dual-channel C2 with the ability to switch between ICMPv4 and DNS tunnelling on the fly. Detection rules using PacketSmith Yara-X modules have been developed to identify traffic generated by this framework.
This content is a critical commentary alleging incompetence among IT personnel in the French public sector, claiming it leads to security risks for citizens. The source is a Reddit post linking to a YouTube video, with no technical details or specific vulnerabilities described. There is no evidence of a concrete security threat, exploit, or vulnerability disclosed in the provided information.
A cyberattack targeted French public websites, including the tax platform, as reported by a French news outlet. The discussion on Reddit highlights concerns about the use of Microsoft products in French government infrastructure. No detailed technical information or specific vulnerabilities have been disclosed.
A publicly accessible GitHub repository allegedly contains data associated with Meckano, an Israeli workforce-management platform. The authenticity, origin, and scope of the data exposure have not been independently verified. The repository's contents require further investigation by security researchers to confirm if this is a genuine data breach. No confirmed exploit or official vendor advisory is available at this time.
Damn Vulnerable NGINX Proxy (DVNP) is a self-contained web security lab published by OWASP that simulates a multi-host NGINX reverse-proxy environment with over 20 real-world misconfigurations. It is designed for security professionals to practice identifying and exploiting chained web and reverse-proxy misconfigurations in a controlled setting. This lab includes backend Flask applications and shared resources to provide a realistic environment for pentesting and bug bounty training. It is not a vulnerability itself but a training tool to improve skills in finding NGINX-related security issues.
This content discusses a security model critique focused on commitment integrity, where an issuer commits to a private artifact by publishing only a public immutable cryptographic fingerprint. The system allows later verification that the artifact has not been altered, addressing threats of an issuer modifying a statement after initial publication. The approach does not guarantee the truthfulness of the statement itself but ensures the integrity of the committed artifact. The discussion includes considerations of key compromise, canonicalization, and the benefits of explicit supersession over simple signed hashes. The source is a Reddit post linking to a demo and security explanation on bestmemecoins.app.
A data breach occurred at ShipMonk, a shipping provider for Trezor, exposing personal details of approximately 13,689 customers who placed orders between May and August 2026. Exposed data includes full names, email addresses, phone numbers, and shipping addresses. Trezor's own systems and hardware wallets were not compromised. The breach increases the risk of phishing attacks targeting affected customers. Trezor has notified impacted individuals and is working with ShipMonk to investigate and secure systems. Customers are advised to be vigilant against phishing and not to share wallet backups or sensitive information.
This content discusses the challenge of accurately determining vulnerable software versions for a given bug. The author created an open source CLI tool called Taggity to automate checking vulnerability descriptions against source code across software releases. The tool helped identify discrepancies in advisories, such as versions incorrectly marked as safe despite lacking fixes. The post is a link to the tool's GitHub repository and does not describe a specific vulnerability or exploit.
This content discusses a security concern related to autonomous AI agents that use large language models (LLMs) for decision-making and execution. It highlights the inherent risks of relying on probabilistic LLM-based guardrails to enforce security policies, which can be manipulated through indirect prompt injection or telemetry poisoning. The proposed solution is to decouple intent generation from execution by implementing deterministic policy gateways that enforce static, immutable authorization checks outside the LLM reasoning loop. This approach aims to prevent adversarial manipulation and unintended self-denial-of-service conditions in AI-driven environments.
This content is a write-up by a security researcher describing their approach to solving a pwn.college MITM challenge. The author overengineered the solution by manually managing TCP connections and sequence numbers instead of using the intended ARP spoofing and transparent forwarding method. The write-up includes technical details about the challenges faced, including an ACK sequence-space bug that disrupted the connection.
This is an open-source, categorized catalog of over 2,800 malware families mapped to frameworks such as NIST CSF, CISA advisories, and MITRE ATT&CK. It is designed to aid incident responders, SOC analysts, and threat hunters by providing a centralized dataset with actionable containment steps and response playbooks. The catalog is freely available under an Apache-2.0 license and accessible via multiple formats including web UI and JSON API. It is not a vulnerability or active threat but a resource for cybersecurity professionals.
This content is a Reddit post where an individual seeks advice on transitioning from a backend development career to offensive cybersecurity. The post includes personal background, career concerns, goals for relocation, and a list of self-study resources. It is not a security threat or vulnerability but rather a career discussion.
Microsoft's August 2026 Patch Tuesday addresses 398 vulnerabilities, including a critical Windows kernel privilege escalation zero-day actively exploited in the wild and four remote code execution flaws with CVSS scores of 9.8. This patch cycle is notable for the volume and severity of fixes, emphasizing the importance of immediate patching to mitigate high-risk vulnerabilities.
A configuration in the Telegram Android app allows app data to be backed up via Android's backup system, potentially exposing sensitive tokens and settings. This backup capability, controlled by the android:allowBackup attribute, can be exploited via adb backup commands if the device is accessible. The exposed data includes authentication tokens and user preferences, which could allow an attacker to hijack sessions without needing email or SMS verification if two-factor authentication is not enabled. The app stores salted and hashed PINs, which could be brute-forced offline. This issue was observed in the app's manifest as of February 3, 2023, and relates to the use of Android's BackupAgentHelper mechanism.
This report analyzes the volume of security fixes in Chrome and Firefox over the last 12 months, highlighting a significant increase in the number of bugs fixed starting in early 2026. The increase correlates with the adoption of AI-assisted vulnerability discovery pipelines, notably Project Glasswing and Anthropic's Claude Mythos Preview model. While the number of CVEs and bugs fixed has surged, this does not necessarily indicate an increase in exploitable vulnerabilities or risk. Other vendors like Apple and Android have not shown similar increases despite access to the same AI models, suggesting that internal pipeline changes drive the volume. The data is based on official Mozilla advisories and CVE records, not press releases.
A survey of 200 US-based CISOs, CTOs, and other senior security leaders at large organizations reveals widespread concern about new security risks introduced by AI deployments. The majority (93%) are worried about these risks, but only a small fraction (15%) are confident their current security tools can handle them. Key issues include a growing attack surface driven by complex non-human identities and insufficient access controls for machine workloads. The survey highlights operational challenges such as lengthy network change management delays and a widening patch gap. Most organizations are actively exploring new approaches to secure non-human identities, as existing identity systems are largely deemed insufficient.
CVE-2026-17106, known as CopyEscape, is a high-severity vulnerability in Docker's 'docker cp' command that allows a malicious container to write arbitrary files to the host filesystem. The flaw stems from a filesystem race condition during archive creation combined with unsafe symbolic link handling during extraction. This can lead to arbitrary file creation or overwriting on the host and potentially code execution depending on the privileges of the Docker CLI user. Docker has released official fixes in Docker Engine/CLI version 29.7.2 and later, Docker Desktop 4.86.0 and later, and Docker Sandboxes 0.38.0 and later.
