Today's briefing shows agentic AI delivering comparable technical capabilities on both sides of the fence, with authorisation as the primary differentiator. Hermes operated unattended in "YOLO" mode against Thailand's Ministry of Finance, enumerating hosts, identifying privilege escalation paths, and cataloguing internal file shares without human approval at each step. On the same day, Kimi K3 agents were credited with discovering 19 Redis zero-days and building an authenticated remote code execution chain in under 30 minutes, while Aikido's pentest agents identified eight high-severity NodeBB vulnerabilities during a six-hour source review. The capability—rapid, autonomous discovery and exploitation of real weaknesses—is increasingly comparable whether exercised by security researchers or attackers; what differs is authorisation and intent.

A separate theme running through today's events is attribution. Google introduced a standardised cryptonym naming system for threat actors to improve consistency in how groups are tracked and discussed. However, better naming does not solve the harder problem of attribution when intrusions — like today's breach of Thailand's Ministry of Finance, where the initial access vector remains unknown — lack sufficient forensic evidence to determine responsibility.

The Clop ransomware gang is exploiting CVE-2026-12569, a critical deserialisation flaw in PTC Windchill and FlexPLM, to deploy JSP webshells and exfiltrate data. Attackers are sending extortion emails from previously compromised accounts to hundreds of employees at targeted organisations. PTC released patches on 17 June, and CISA has added the flaw to its Known Exploited Vulnerabilities catalogue.

Attackers compromised a maintainer's GitHub account to publish mrmustard 0.7.4, a poisoned version of Xanadu's quantum computing library, directly to PyPI. On import, the payload steals SSH keys, AWS credentials and Kubernetes configs, then installs cron, shell and Python startup persistence mechanisms. The malicious code targeted research and HPC environments, also collecting SLURM job data and GPU inventories before exfiltrating everything to an attacker-controlled server.

Europol supported a multi-country operation in June and July 2026 against online content linked to "The Com," a loosely organised nihilistic extremist network. Investigators from nine countries referred roughly 4,340 URLs for removal, targeting propaganda, grooming material, and content depicting self-harm and violence aimed at minors. The action forms part of Project COMPASS, coordinated by Europol's European Counter Terrorism Centre.

Researchers found an autonomous AI agent, Hermes, running in unattended "YOLO" mode to enumerate Thailand's Ministry of Finance network via exposed staging server directories. The operation staged a custom Go implant named Hades, exploit code for multiple CVEs, and tools targeting Hadoop, mail, and web infrastructure. Agent logs showed privilege escalation scanning and file enumeration, though no evidence confirmed data exfiltration.

Redis released seven security updates on 23 July after researchers published proof-of-concept exploits for authenticated remote code execution across several versions. The flaws stem from a Streams use-after-free bug and a RedisBloom TDigest out-of-bounds write, both reachable via the RESTORE command. Researchers using Kimi K3 AI agents reportedly discovered the vulnerabilities, though Redis has not assigned new CVE identifiers and no in-the-wild exploitation has been confirmed.

A site at corepack.org impersonates the Node.js Corepack tool to distribute an infostealer and proxyware to developers seeking downloads. Clicking through delivers a fake VPN installer that harvests browser data and SSH keys while enrolling machines as bandwidth-sharing proxy nodes. The site exploits Corepack's removal from newer Node.js versions and ranks highly in search results for related queries.

Google Threat Intelligence Group is introducing a unified naming schema, merging former Mandiant and TAG tracking systems into memorable two-word cryptonyms. The first word identifies the specific actor, while the second denotes origin or motivation, such as CASTLE for China or COMET for cybercriminal groups. Previous names remain searchable within the Google Threat Intelligence platform alongside MITRE ATT&CK mappings.