In this week's Seclog, the cybersecurity landscape is markedly shaped by the rapid evolution of AI, both as a tool for attackers and a subject of critical safety research. We see new vulnerabilities emerging in AI-driven systems, from data exfiltration in Google's Gemini to RCE in the Antigravity IDE, alongside the alarming rise of AI/LLM-generated malware. Furthermore, the ethical implications of AI's use in bug bounty platforms sparked significant debate, highlighting concerns over intellectual property and trust. Traditional attack vectors remain prevalent, with critical RCEs impacting widely used software like BeyondTrust and SmarterMail, while novel exploitation techniques leveraging HTTP trailer parsing discrepancies and HMAC collisions demonstrate ongoing innovation from adversaries. The release of advanced offensive tools for SSRF, template injection, and Kerberos attacks, alongside defensive resources for Azure attack paths and spying browser extensions, underscores the continuous cat-and-mouse game between offense and defense. Overall, the content emphasizes the growing complexity of securing modern environments, particularly with the integration of increasingly autonomous and powerful AI technologies.

📚 SecMisc #

ReMemory: Shamir's Secret Sharing - eljojo.github.io

Introduces ReMemory, a tool that encrypts files and distributes the decryption key using Shamir's Secret Sharing algorithm.
Enables users to specify a threshold of trusted individuals required to reconstruct the key, ensuring no single person can unilaterally access sensitive data.
Highlights its robust design for offline use and self-contained recovery bundles, ensuring data accessibility even if the host website is unavailable.

📰 SecLinks #

Enigma SSRF: Autonomous Fuzzer for Server-Side Request Forgery - labs.trace37.com

Introduces Enigma, an adaptive SSRF engine designed to automate the discovery and exploitation of Server-Side Request Forgery vulnerabilities.
Highlights advanced techniques used by Enigma, including IP obfuscation, URL parser confusion, and protocol smuggling, which can bypass common SSRF defenses.
Emphasizes the use of out-of-band (OOB) callback confirmation to reliably validate SSRF exploits, crucial for automated scanning and identifying blind SSRF scenarios.

PhoneLeak: Gemini Data Exfiltration via Phone Call - blog.starstrike.ai

Details a novel data exfiltration vulnerability discovered in Google's Gemini AI, leveraging phone call functionalities.
Highlights an unusual attack vector, demonstrating how sensitive data can be siphoned off through non-traditional communication channels within AI systems.
Emphasizes the importance of scrutinizing all output and communication methods of complex AI models, not just their direct text or API interactions, for potential leakage points.

Age Verification Bypass Tool Analysis - age-verifier.kibty.town

Describes a method to bypass age verification systems used by platforms like Discord, Twitch, and Snapchat, leveraging a client-side script injection.
Explains that the bypass exploits the metadata-based verification process of providers like k-id, rather than facial recognition images, allowing crafted data to simulate legitimate verification.
Highlights a potential privacy-security trade-off, where a system designed for user privacy (not sending raw images) becomes vulnerable to metadata manipulation for fraudulent age verification.

BeyondTrust RCE: Reconnaissance Activity Detected - greynoise.io

Reports the detection of active reconnaissance targeting a critical pre-authentication Remote Code Execution (RCE) vulnerability, CVE-2026-1731, in BeyondTrust Remote Support and Privileged Remote Access.
Indicates that a proof-of-concept (PoC) exploit for this high-severity vulnerability (CVSS 10/10) was publicly released on GitHub, triggering immediate scanning activity.
Urges organizations using BeyondTrust products to apply patches immediately and monitor for compromise, as attackers are actively searching for vulnerable instances.

Pharmacy Hack: Insecure Admin APIs Exploited - eaton-works.com

Details the exploitation of insecure super admin APIs on Dava India Pharmacy's website, allowing for the creation of a high-privileged account.
Explains that this access granted full control over the pharmacy backend, including customer orders, product details (allowing price changes and removal of prescription requirements), and coupon generation.
Underscores the critical impact of poorly secured API endpoints, which can lead to extensive data breaches, financial fraud, and compromise of critical business operations.

AI Agent Intelligence Analysis Platform - blog.lukaszolejnik.com

Introduces ClawdINT, an experimental intelligence analysis platform designed for AI agents to operate as first-class users.
Describes its purpose as enabling AI agents to autonomously register, research current events, and publish structured assessments across various domains like cybersecurity and geopolitics.
Explores the emerging paradigm of collaborative AI platforms where agents can independently gather, analyze, and disseminate intelligence, posing both opportunities and new security considerations for agent interactions and data integrity.

Firefox RCE: Typo in Wasm Component - kqx.io

Details how a seemingly simple typo within the SpiderMonkey Wasm component of Firefox led to a critical Remote Code Execution (RCE) vulnerability.
Highlights the profound impact that subtle coding errors can have in complex software, enabling attackers to gain control over user systems.
Underscores the importance of meticulous code review, advanced static analysis, and fuzzing to detect even minor flaws that could become critical security issues in widely used applications.

CodeThreat AI Hub: AI Component Vulnerabilities - hub.codethreat.com

Introduces the CodeThreat AI Hub, a platform offering security intelligence specifically for the AI component ecosystem.
Provides open vulnerability data and insights relevant to Machine Learning, Container, and Platform (MCP) servers and AI Agent Skills.
Addresses the emerging need for specialized security analysis tools focused on the unique attack surface and potential vulnerabilities within AI-driven systems.

HTTP Trailer Parsing Discrepancies - sebsrt.xyz

Explores the security implications arising from inconsistent handling of HTTP/1.1 trailer fields across different HTTP implementations.
Identifies potential attack vectors where varied interpretations of these rarely used headers can lead to bypasses, information disclosure, or other vulnerabilities.
Advises security professionals to be aware of how their applications and infrastructure process HTTP trailers, as discrepancies can create unexpected weak points.

Google Antigravity RCE Achieved - hacktron.ai

Details the discovery and exploitation of a Remote Code Execution (RCE) vulnerability in Google's new AI code editor, Antigravity.
Highlights that Antigravity shares underlying mechanisms with the previously known Windsurf IDE, suggesting potential reuse of vulnerable components or design patterns.
Emphasizes the critical security risks associated with cloud-based IDEs, as RCE flaws can grant attackers extensive access to development environments and sensitive codebases.

HMAC Collisions Forge Password Tokens - asdf.foo

Describes a technique leveraging HMAC collisions to forge password reset tokens, potentially allowing an attacker to change a victim's account password.
Explains how subtle weaknesses in HMAC implementation or secret management can be exploited to generate valid-looking tokens without access to the original secret.
Underscores the importance of robust cryptographic practices for sensitive tokens and the need for careful review of password reset mechanisms to prevent such account takeover vulnerabilities.

AI-Generated Malware Exploits React2Shell - darktrace.com

Reports the discovery of malware generated by AI/LLMs actively exploiting the React2Shell vulnerability in a cloud environment.
Illustrates a concerning trend where AI-assisted development lowers the barrier for entry for low-skill attackers, enabling them to rapidly create effective exploitation tools.
Highlights the increasing challenges for defenders in detecting and responding to AI-generated threats, necessitating advanced behavioral analysis and threat intelligence to keep pace with evolving attack methodologies.

International AI Safety Report 2026 - internationalaisafetyreport.org

Presents the second International AI Safety Report (2026), a comprehensive review of the capabilities and risks of general-purpose AI systems.
Emphasizes the collaborative nature of the report, involving over 100 AI experts from 30+ countries and international organizations, signifying a global consensus on AI safety importance.
Offers critical insights for policymakers, researchers, and security professionals on the evolving threat landscape posed by advanced AI, guiding future regulatory and defensive strategies.

YAML Parser Differential Vulnerabilities - blog.darkforge.io

Explores "parser differential" vulnerabilities, specifically focusing on how YAML files can be interpreted differently by various parsers.
Demonstrates how these discrepancies can be exploited to create security flaws, potentially leading to unexpected code execution or data manipulation depending on the parsing engine.
Provides new techniques for crafting YAML payloads that can confuse multiple parsers without relying on binary tags, underscoring the subtle complexities in data serialization formats that attackers can leverage.

Augustus: LLM Prompt Injection Tool - praetorian.com

Introduces Augustus, an open-source tool designed for testing Large Language Model (LLM) services for prompt injection vulnerabilities.
Functions as a follow-up to Praetorian's Julius tool, which identifies the underlying LLM infrastructure, enabling a comprehensive assessment workflow.
Empowers security professionals to evaluate the security posture of LLM deployments by actively probing for weaknesses that could lead to unauthorized actions or data exfiltration via crafted prompts.

Exploring Azure Attack Paths - cloudbrothers.info

Provides insights into various attack paths and common misconfigurations within Microsoft Azure cloud environments.
Emphasizes the increasing complexity of securing Azure due to its rapidly expanding services and features, highlighting the need for continuous vigilance.
Offers guidance on identifying bad practices to avoid and understanding prevalent attack scenarios to help organizations fortify their cloud security posture.

Perplexity Comet: Agentic Browser Reversing - labs.zenity.io

Presents a technical deep dive into Perplexity's Comet, an "agentic browser" that allows an AI model to autonomously interact with web pages.
Dissects the architectural design, detailing the communication mechanisms between the AI model and the browser, as well as the tools available to the model.
Provides critical insights into how AI agents perceive and interact with web content, revealing potential new attack surfaces and defensive challenges in the realm of autonomous AI web agents.

SmarterMail Pre-Auth RCE (CVE-2025-52691) - labs.watchtowr.com

Details a critical pre-authentication Remote Code Execution (RCE) vulnerability (CVE-2025-52691) in SmarterTools' SmarterMail solution, rated 10/10 CVSS.
Highlights the severe impact of this flaw, allowing attackers to achieve full system compromise without prior authentication.
Urges immediate patching and thorough security audits for organizations utilizing SmarterMail, as pre-auth RCEs are prime targets for opportunistic exploitation.

🐦 SecX #

HackerOne AI Usage Concerns - x.com

Expresses significant concerns from the bug bounty community regarding platforms like HackerOne potentially using submitted vulnerability reports to train AI models.
Highlights accusations that leveraging bug hunter work for AI profit constitutes "stealing research" and breaches client agreements regarding data ownership.
Raises ethical and trust issues within the bug bounty ecosystem, potentially leading to a re-evaluation of how researchers share sensitive vulnerability data with intermediaries.

LLMs for Vulnerability Prioritization - x.com

Acknowledges the potential of Agentic Large Language Models (LLMs) to automate vulnerability detection, but critically points out their current weakness in vulnerability prioritization.
Emphasizes that the real challenge in vulnerability research lies in effectively exploring the search space and distinguishing critical signal from irrelevant noise.
Mentions the development of a paper and an open-source tool designed to address this prioritization gap, aiming to improve the efficiency and impact of LLM-assisted vulnerability analysis.

🎥 SecVideo #

Image Disguise: RAT Payload Delivery - youtube.com

Discusses a technique where malicious Remote Access Trojans (RATs) are disguised as benign image files, enabling stealthy initial access.
Implies a deep dive into the methods used by attackers to embed and execute malware from seemingly innocuous file types, circumventing traditional file-type-based defenses.
Suggests the need for advanced detection mechanisms beyond file extensions, such as behavioral analysis or content inspection, to identify such threats.

💻 SecGit #

Pirebok: Evolutionary Guided Adversarial Fuzzer - github.com

Introduces Pirebok, an open-source guided adversarial fuzzer leveraging evolutionary search algorithms.
Indicates its utility for discovering complex vulnerabilities by intelligently exploring input spaces beyond traditional brute-force methods.
Suggests its application in security testing to uncover edge cases and subtle flaws in software, particularly where sophisticated input generation is required.

Report on Spying Browser Extensions - github.com

Provides a report by the Q Continuum group detailing various browser extensions identified as performing surveillance or data collection beyond their stated purpose.
Offers insights into the methods and data points collected by these malicious extensions, highlighting risks to user privacy and enterprise data.
Serves as a resource for security professionals to identify and mitigate risks associated with untrusted or compromised browser extensions in their environments.

Template Injection Testing Playground - github.com

Presents an open-source "Template Injection Playground" designed to facilitate testing for server-side template injection (SSTI) vulnerabilities across numerous template engines.
Offers a practical environment for security researchers and developers to understand and identify potential injection points in applications leveraging various templating technologies.
Provides a valuable resource for both red team operations to discover vulnerabilities and blue team operations to validate their defenses against SSTI attacks.

AutoPtT: Kerberos Pass-the-Ticket Tool - github.com

Introduces AutoPtT, a standalone tool designed for automating Kerberos Pass-the-Ticket (PtT) attacks.
Offers an alternative to established tools like Rubeus or Mimikatz, implemented in C++ and Python for interactive or step-by-step execution.
Provides red teamers with a versatile utility for post-exploitation lateral movement within Active Directory environments by exploiting Kerberos authentication mechanisms.

← Back to Seclog