Seclog - #187

In this week's Seclog, the dynamic interplay between artificial intelligence and cybersecurity takes center stage, illustrating both the escalating capabilities of AI in offensive security and the innovative defensive countermeasures emerging in response. We see AI-driven approaches to vulnerability discovery, ranging from autonomous voltage fault injection to bypass secure boot, to leveraging advanced LLMs for uncovering high-value RCEs in widely deployed software like WordPress with minimal effort. Simultaneously, the community is developing novel AI defenses, including context bombs designed to neutralize AI attackers and canary tokens for effective prompt injection detection, highlighting the ongoing arms race in securing AI systems. Beyond AI, significant vulnerabilities continue to surface, notably a 15-year-old pre-authentication RCE in Nginx, and new methods for macOS password hash dumping and ServiceNow public data exposure are revealed. The importance of robust software supply chain security is also underscored, alongside updates to essential tools for static analysis and .NET deserialization payload generation, collectively showcasing a vibrant and rapidly evolving security landscape.

📰 SecLinks #

Context Bombs Defeat AI Attackers - agentic.tracebit.com

This technique introduces "context bombs" as a proactive defense mechanism against adversarial AI attacks. It focuses on embedding specific patterns within the AI model's context to disrupt or neutralize malicious prompts and agentic behaviors.

Supply Chain Security Manifesto Released - supplychain.exposed

This manifesto advocates for a fundamental shift in software supply chain security, emphasizing pre-publication scanning by registries. It highlights the critical role of package managers as authoritative sources and calls for free, enriched, and high-quality vulnerability data. The core thesis suggests that robust detection mechanisms throughout the supply chain are the ultimate control for preventing malware ingress.

EthiBench Evaluates Real-World AI Pentesting - ethiack.com

Traditional pentesting benchmarks are inadequate for assessing AI agent capabilities in realistic scenarios. Ethiack developed EthiBench, a novel evaluation framework designed to test AI pentesting agents against practical targets. EthiBench incorporates structured ground-truth, LLM-as-a-judge methodologies, and metrics extending beyond simple flag capture to provide a more comprehensive assessment.

Semgrep C/C++ Ruleset Updated - hnsecurity.it

An enhanced Semgrep ruleset for C/C++ vulnerability research has been significantly upgraded. This update delivers improved accuracy and performance for static analysis of C/C++ codebases. The ruleset is now publicly available in the Semgrep registry, offering a valuable resource for identifying common and complex vulnerabilities.

macOS Password Hash Dumping Guide - labs.sra.io

This guide provides methods for extracting local user hashes from macOS environments, a common blind spot in corporate security. It details the process of transforming these hashes into a format compatible with tools like Hashcat for offline cracking. Security professionals can use this information to assess the susceptibility of macOS endpoints to credential compromise through direct access or post-exploitation scenarios.

Snowpick Reveals ServiceNow Data Exposure - bishopfox.com

Bishop Fox developed "snowpick," a new tool designed to identify public data exposure within ServiceNow instances. The tool specifically targets vulnerabilities in ServiceNow widgets and the Table REST API, often overlooked vectors for data leakage. Initial assessments revealed significant findings, with 31% of 166 tested instances exhibiting public data exposure, highlighting a pervasive configuration risk.

AI Bypasses Secure Boot via Fault Injection - raelize.com

Researchers demonstrated that Claude Code can autonomously conduct voltage fault-injection campaigns to bypass secure boot mechanisms. Using a ChipWhisperer Husky, the AI successfully exploited ESP32 Secure Boot V1 by independently developing and executing its own tooling. This research highlights the emerging threat of AI-driven hardware attacks, capable of intricate, hands-off exploitation of embedded systems.

WP2Shell WordPress RCE Deep Dive - blog.zsec.uk

This article offers a detailed code trace and deep dive into the wp2shell exploit, which targets WordPress for Remote Code Execution (RCE). It explains the technical specifics of the vulnerability, including how the RCE is achieved through specific WordPress components. Security professionals can gain a deeper understanding of the attack vector and the underlying mechanisms, aiding in defensive strategies and vulnerability analysis.

Nginx Pre-Auth RCE Discovered - cyberstan.co.uk

A 15-year-old pre-authentication Remote Code Execution (RCE) vulnerability has been discovered in Nginx, affecting numerous call sites. The vulnerability, identified as CVE-2026-42533, leverages a "two-pass capture clobbering" technique. This critical RCE allows unauthenticated attackers to execute arbitrary code, posing a significant threat to Nginx deployments and requiring immediate patching.

AI Discovers WordPress RCE for $25 - slcyber.io

This research demonstrates the efficacy of AI, specifically GPT5.6 Sol Ultra, in discovering high-value WordPress Remote Code Execution (RCE) vulnerabilities. The RCE was found with minimal investment ($25), highlighting the increasing accessibility and power of AI for vulnerability research. A tool, wp2shell.com, is provided for defenders to check their WordPress instances against this newly disclosed vulnerability, emphasizing responsible disclosure practices.

LLM Prompt Injection Detection with Canaries - toxsec.com

This article introduces "canary tokens" as an economical and effective method for detecting prompt injection attempts in Large Language Model (LLM) applications. The technique involves embedding high-entropy strings within the LLM's context and monitoring the output for their unauthorized extraction. Successful extraction of these tokens acts as a tripwire, immediately indicating a prompt injection attempt and signaling potential compromise or data exfiltration.

Demystifying Kimi K3: China's 2.8T Parameter Frontier Model - substack.com

Technical deep-dive into Moonshot AI’s 2.8-trillion-parameter open-weight sparse Mixture-of-Experts model Kimi K3. Explores key innovations such as Kimi Delta Attention, Attention Residuals, and Stable LatentMoE that enable a 1M token context window and its #1 ranking on the Frontend Code Arena.

💻 SecGit #

.NET Deserialization Payload Generator - github.com

This GitHub repository provides "ysonet," a robust tool for generating deserialization payloads across multiple .NET formatters. Security researchers and penetration testers can leverage this tool to identify and exploit deserialization vulnerabilities in .NET applications. The utility supports various common .NET serialization formats, simplifying the process of crafting malicious payloads for testing.

Tailscale Mullvad Toggle Tool - github.com

No-cloud, no-VPS solution for easily toggling Mullvad exit nodes in Tailscale. Provides a dependency-free ts-mullvad CLI along with a macOS SwiftBar menu bar integration. Supports country/city-based toggling, status checking, and public egress verification.