NGINX HTTP/3 zero-day, FortiBleed VPN hijack, Apple SecureROM cracked

A newly disclosed use‑after‑free in nginx’s ngx_http_v3_module lets a remote, unauthenticated client reopen a QPACK encoder stream, corrupting freed memory. Exploitation can cause worker crashes and, if memory layout is favorable, remote code execution. The flaw affects any nginx build with HTTP/3 enabled and is scored 9.2 CVSS.

The UnCanny GitHub proof‑of‑concept shows a new LPE primitive where a standard user supplies a UNC path in the Windows Store InstallService’s plugin JSON, causing the system service (InstallService.exe) to load the path and issue outbound NTLM as the machine account. This reveals a machine‑account NTLM coercion vector that can be leveraged for privilege escalation on unpatched Windows machines.

Paradigm Shift released usbliter8, a proof‑of‑concept that achieves arbitrary code execution in the SecureROM of Apple A12 and A13 chips, a hardware bug that cannot be patched via software. The flaw compromises the full boot‑chain, exposing devices to persistent takeover, and the only realistic mitigation is hardware replacement.

ESET’s 2026 analysis reveals the ransomware‑as‑a‑service gang Gentlemen runs a home‑grown ‘GentleKiller’ suite that can disable endpoint detection and response tools within hours of a vulnerability’s disclosure. The framework integrates third‑party killers and spoofed vendor signatures, widening the gang’s reach beyond the US. Defenders must watch for its driver‑based evasion tricks.

Researchers uncovered a coordinated operation that launched 1.16 billion login attempts against 320 k FortiGate appliances, cracked NTLM/Kerberos hashes on a 45‑GPU RTX 4090 cluster, and replayed VPN cookies to infiltrate Active Directory. The scale shows how attackers can turn credential‑spraying into full network compromises, forcing defenders to rethink VPN exposure.

Palo Alto Networks identifies seven ways attackers can corrupt AWS CloudTrail and Google Cloud Logging, either to hide their activity or siphon logs for continuous surveillance. The techniques prove that compromising logging services can give threat actors persistent, undetected access, forcing defenders to harden the very tools they rely on for detection.

Klue’s Battlecards integration was compromised, allowing the Icarus group to harvest OAuth tokens and query connected Salesforce instances. The attackers exfiltrated CRM records and then launched extortion campaigns against the affected customers, prompting Salesforce to disable the Klue app. This supply‑chain breach highlights the risks of third‑party SaaS connections.

Anthropic's Fable model was shut down after the U.S. classified it as a dual‑use munition, forcing a blanket export ban. The episode shows how quickly AI capabilities can outpace existing export‑control frameworks, exposing a regulatory gap that could hinder both security and innovation.

The Justice Department and Homeland Security seized CFAKE.com and SOCFAKE.com, which hosted thousands of non‑consensual AI‑generated nude images of public figures. The move marks the first federal action under the 2025 TAKE IT DOWN Act, signaling tougher enforcement against deep‑fake pornography and setting a precedent for protecting victims’ privacy.