Major Incidents or Breaches

  • Three researchers at Hacktron used Anthropic’s Claude Opus 5 to chain two vulnerabilities, allowing them to take over ChatGPT and Codex accounts belonging to several OpenAI employees, ultimately gaining access to internal OpenAI systems. The attack leveraged Claude Opus 5 to facilitate the exploitation process [1].
  • Researchers successfully escaped the OpenAI Codex sandbox using two separate methods, one of which enabled them to execute commands on a developer’s machine even in the most restricted mode. OpenAI has since patched both vulnerabilities [4].
  • The ShinyHunters extortion group breached the Clop ransomware operation’s data leak site, defaced its Tor site, and reportedly stole server data and private keys for the onion service, threatening to extort the Clop gang [7].
  • The North Korean WaterPlum threat group compromised at least 30,000 devices globally between December 2025 and July 2026, transferring over $10.7 million in illicit funds. The campaign was disclosed in a joint law enforcement advisory [6].

Newly Discovered Vulnerabilities

  • Security researchers demonstrated that AI browser agents including those in Chrome, Edge, Opera Neon, Perplexity Comet, and Claude in Chrome can be hijacked through a malicious extension using the BragJack proof-of-concept attack. The attack allows full control over the AI assistant within the browser [5].

Notable Threat Actor Activity

  • The WaterPlum group, attributed to North Korea, conducted widespread infections and financial theft via a global campaign affecting tens of thousands of devices [6].
  • ShinyHunters, typically known for data extortion, targeted another cybercriminal group by breaching the Clop ransomware gang’s infrastructure and threatening further extortion [7].

Trends, Tools, or Tactics of Interest

  • Chaining of vulnerabilities involving AI models and browser assistants has enabled new attack paths, as demonstrated by both the Claude Opus 5 incident and the BragJack attack [1][5].
  • Exploitation of sandbox escape techniques remains a significant risk in AI development environments, as evidenced by the Codex sandbox escape [4].
  • Use of malicious browser extensions as a vector for hijacking AI assistants is an emerging threat highlighted by the BragJack proof-of-concept [5].

Sources#

  1. Claude Opus 5 Helped Researchers Take Over OpenAI Staff Accounts via Chained Flaws - thehackernews.com
  2. Researchers escape OpenAI Codex sandbox to run commands on host - bleepingcomputer.com
  3. ShinyHunters hacks Clop leak site, threatens to extort ransomware gang - bleepingcomputer.com
  4. North Korean WaterPlum hackers infected 30,000 devices worldwide - bleepingcomputer.com
  5. BragJack attacks hijack AI browser agents through malicious extensions - bleepingcomputer.com