Parse notes
Sponsored
Engineering and security teams often rely on tools that reveal only a narrow slice of what is happening inside their software. The deeper risks in modern code such as logic failures, unexpected execution paths, and behavioral deviations from intent rarely appear without more advanced analysis.
Fortinet has patched CVE-2026-21643, a critical SQL injection vulnerability in the FortiClientEMS GUI that allows unauthenticated remote attackers to execute arbitrary code or commands by improperly neutralizing special characters in SQL queries. The flaw affects FortiClientEMS version 7.4.4. Version 7.4.5 was released as a fix. Versions 7.2 and 8.0 are unaffected. While no exploitation in the wild or public proof-of-concept has been observed yet, organizations should patch urgently, given Fortinet's history as a heavily targeted vendor per CISA's Known Exploited Vulnerabilities Catalog.
Forcepoint X-Labs discovered the Phorpiex botnet delivering Global Group ransomware, a Mamona successor, via phishing emails containing .lnk files disguised as documents. It leverages living-off-the-land techniques using PowerShell and Command Prompt for payload delivery. The ransomware operates in a mute mode that generates encryption keys locally with ChaCha20-Poly1305 instead of contacting a C2 server, allowing it to encrypt files on offline machines and evade network-based detection. After encryption, the malware self-deletes, destroys Volume Shadow Copies, and appends a .reco extension to locked files, leaving minimal forensic artifacts.
Researchers at the security firm LayerX uncovered a critical (CVSS 10/10) vulnerability in Claude Desktop Extensions (DXT) that could enable remote code execution (RCE) with no user interaction. In the proof-of-concept disclosed by LayerX, an attacker would send the victim a Google Calendar invite containing malicious instructions in the event description. The victim would then, at some point, ask Claude to manage their calendar or check events, and the malicious instructions would be executed. This vulnerability is enabled by the fact that Claude DXT intentionally runs without any isolation.
A security researcher discovered that the DoJ's latest Epstein document dump inadvertently included unredacted email attachments as raw base64-encoded content — left uncensored because the assigned intern didn't recognize the significance of pages of encoded data embedded in printed-and-scanned emails. Reconstructing the original PDFs proved extremely difficult due to poor OCR quality on Courier New font rendering (where "1" and "l" are nearly indistinguishable), with multiple commercial and open-source OCR tools failing until a community-developed pixel-level template-matching approach achieved 99.96% line accuracy and successfully decompressed 39 of 40 FlateDecode streams. The incident highlights how encoding artifacts and document processing pipelines can inadvertently preserve sensitive content that redaction workflows miss, and serves as a cautionary tale for forensic document handling and redaction verification processes.
In November 2025, the Sysdig Threat Research Team (TRT) detected an account compromise that was notable for the attacker's 8-minute progression from initial access to administrative privileges, using LLM assistance. The attacker gained initial access through credentials discovered in a public S3 bucket. They then used these credentials to replace the Lambda function code to escalate privileges, move laterally across 19 different principals, and finally harvest sensitive data. They also used Bedrock to access cloud-hosted LLMs and launch GPU instances for model training. The article includes mitigation recommendations for each step in the attacker's exploit chain.
Sponsored
AI lets developers ship in seconds, and now security can move just as fast. depthfirst [11] understands your code, business logic, and infrastructure to trace full attack paths, validate what's actually exploitable, and tell your devs exactly how to fix it. Trusted by Lovable, Supabase, Moveworks, and Persona. Book a demo and try it for
Augustus is an open-source Go-based LLM vulnerability scanner that tests models against 210+ adversarial attacks across 47 categories, including jailbreaks, prompt injection, encoding exploits, data extraction, and agent manipulation, shipping as a single binary with support for 28 LLM providers. Inspired by NVIDIA's garak but reimplemented in Go for faster execution and zero runtime dependencies, Augustus features a pipeline architecture with composable probes, buff transformations (encoding, paraphrase, poetry, and low-resource language translation), and 90+ detectors, including LLM-as-a-judge and HarmJudge. The tool is part of Praetorian's "The 12 Caesars" campaign.
Klint is a single-binary, zero-dependency incident response scanner written in C++23 that detects hidden kernel modules, rootkits, compromised syscall tables, and eBPF-based persistence mechanisms across 8 self-registering scanners. The tool cross-references multiple kernel data sources — including /proc/kcore, /proc/kallsyms, MSRs, and cgroup task files — to identify inconsistencies indicative of tampering, with structured JSON output and meaningful exit codes designed for CI pipeline integration. Builds as a fully static binary for rapid deployment via SCP into compromised environments during live incident response.
Leash wraps AI coding agents in containers, monitors their activity, and enforces custom Cedar policies instantly.
Germany's BSI and BfV warned that state-backed hackers are targeting military leaders, diplomats, and journalists across Europe by abusing Signal's legitimate features — impersonating Signal Security Support to steal PINs and hijack accounts, or tricking targets into scanning QR codes that silently link attacker devices with access to 45 days of chat history. The campaigns rely entirely on social engineering rather than exploits, making them difficult to detect through traditional security tooling. Users should enable Registration Lock, audit linked devices regularly, and treat any unsolicited messages requesting verification codes as malicious.
A hacktivist going by the name wikkid has scraped and released over 500K customer records from stalkerware services provided by a Ukrainian company called Struktura. The scraped data includes customer email addresses, which app or brand the customer paid for, how much they paid, the payment type, and the last four digits of credit card numbers. The hacktivist stated that they exploited a “trivial bug” in the service's website.
Discord has announced that beginning in early March, users will need to verify their age via a selfie or government ID. Users' accounts will be set to teen accounts by default until they perform the verification. Discord states that the selfie processing will happen locally on users' devices, that government IDs will be deleted by their third-party processors “quickly,” and that age verification status will not be visible on users' profiles.
Sponsored
AuditBoard's free virtual event features risk experts from EY, Kimberly-Clark, and OpenAI. Get up to 4 free CPE credits as you learn
Two Connecticut men were indicted on 45 federal counts for allegedly using ~3,000 stolen identities purchased from darknet markets to create fraudulent accounts on FanDuel, DraftKings, and BetMGM, exploiting new-user promotional bonuses to steal $3 million over five years.
Google Research used Gemini's iterative self-correction prompting to discover a bug in a cryptography paper on SNARGs for NP from LWE that human reviewers had missed, highlighting LLMs' growing potential as adversarial reviewers in academic research and security audits.
Microsoft announced smartphone-style app permission prompts and a "Windows Baseline Security Mode" for Windows 11.