Thermo Fisher patched a flaw (CVE‑2026‑17583) that let attackers modify .fsa and .hid DNA files without detection, potentially spoofing forensic results. The update adds mandatory digital signatures for five active product lines, forcing labs to verify file integrity before analysis.
Unit 42 discovered that many passwordless systems, especially Google’s synced passkeys, skip verification of the User‑Verified flag. Malware on a compromised device can then forge a passkey authentication, effectively reducing multi‑factor authentication to a single factor and hijacking accounts. The finding forces defenders to audit their passkey validation logic.
Truffle Security scanned every public Hugging Face dataset, 7.6 petabytes across 187 million files, and uncovered 221,303 unique live credentials in 6,003 datasets. Among them were 349 GitHub personal access tokens with repo‑write scope and thousands of AWS and Hugging Face API keys, giving attackers the ability to inject malicious code into widely‑used AI models.
AI‑driven tools are flooding the National Vulnerability Database with hundreds of low‑quality CVE entries, forcing NIST to halt enrichment for all but a few high‑risk categories. The backlog now exceeds 30,000 CVEs, meaning many real threats may go untracked and defenses lose a reliable source of truth.
A CVE‑2026‑4810 flaw in Google’s Agent Development Kit lets an attacker embed malicious prompts in a pull request, causing one AI agent to hijack another and execute arbitrary code. The bug, rated critical, was patched in ADK 1.28.1 and 2.0.0a2.
INC ransomware has hijacked the newly disclosed SonicWall SMA‑1000 VPN flaws (CVE‑2026‑15409/10), accelerating attacks to 885 victims across government and private sectors. The chain enables credential theft, MFA seed extraction, and deep lateral movement, forcing organizations worldwide to patch immediately or face persistent compromise.
DOUBLECUP, a Russian loader‑as‑a‑service, hides malicious code in PNG images cached by browsers via ClickFix prompts, then steals the payload with certutil or findstr and runs it in memory. It delivers the CountLoader dropper to Windows/macOS and a new Windows RAT called DeviceManager, expanding the attack surface for low‑skill operators.
A deep reverse‑engineering teardown maps csagent.sys’s six kernel callbacks, process, thread, image, object, registry, and file, plus a WFP network engine that can inject TCP resets to kill C2. The reveal shows exactly how Falcon monitors endpoints and where attackers might slip past its detection.
The repo provides a set of PowerShell tools that impersonate sandboxed process tokens to probe what resources a sandbox can access, letting researchers spot misconfigurations. It includes utilities for token inspection, memory section handling, and NT object enumeration, and runs best with admin rights.
Back Engineering Labs reverse‑engineered Tencent’s ACE anti‑cheat virtual machine, exposing a static devirtualization that translates its .tvm dispatcher while preserving CET shadow‑stack and SEH semantics. The analysis shows the VM’s boxed native instructions and RDSSPQ/INCSSPQ balancing are vulnerable to guided symbolic evaluation, making cheat‑bypass on Linux/Proton feasible.
You’re caught up.
That was all 10 stories in the 3 August issue.