Investigators Trace AI-Assisted Bank Heists to TAG-217 Group Using Claude, DeepSeek
Researchers link South Korean bank breaches to a suspect using multiple AI models and open-source testing frameworks to scan for vulnerabilities.
Key points
- The TAG-217 group used Claude Code, DeepSeek, Grok, and Zhipu AI alongside the ARTEX framework during attacks on South Korean financial institutions.
- Public reporting indicates personal data belonging to at least 68,000 individuals across seven banks was exposed in these incidents.
- Recorded Future’s Insikt Group identified the actor as a suspected Chinese speaker, though state backing has not been confirmed.
Researchers from Recorded Future’s Insikt Group have linked a series of cyberattacks on South Korean banks to a threat actor designated as TAG-217. The investigation reveals that this group leveraged multiple artificial intelligence services, including Claude Code, DeepSeek, and Grok, to assist in identifying system weaknesses.
The campaign combined these AI tools with ARTEX, an open-source penetration testing framework, to probe various financial systems. According to GBHackers, this approach demonstrates how attackers integrate generative AI with established hacking methodologies to search for vulnerabilities across different infrastructure components while leaving behind traceable digital evidence.
Operational Timeline and Tactics
In a report published on October 16, Insikt Group detailed how TAG-217 began preparations in August, weeks before the banking breaches became public knowledge. The team tracked the actor’s activities between August 17 and August 22, noting tests on remote-access infrastructure and scans of internet-facing systems.
The investigators found logs showing repeated connections between the attacker’s own servers during this period. However, these records did not prove successful access to victim networks at that stage. Consequently, these early activities should be viewed as preparatory testing rather than confirmed breaches of the financial institutions.
Attribution and Evidence
Recorded Future described the operator as a suspected Chinese speaker but stopped short of labeling the group as state-sponsored. The attribution relies on technical artifacts left behind, including exposed Claude Code sessions, ARTEX configuration settings, and AI memory files found on compromised systems.
Public analysis on October 7 identified the specific AI tools used in the workflow. While these records clarify the technical process, the researchers noted that available evidence cannot quantify how much the AI accelerated the attack timeline. The findings also do not establish that the AI tools operated without direct human guidance.
Impact on Financial Institutions
Public reporting cited in the investigation states that the breaches exposed personal information belonging to at least 68,000 people. The affected entities include at least seven distinct financial institutions within South Korea.
The wider investigation into the South Korean banking sector followed disclosures involving different services and types of data. The researchers noted that there was no single shared weakness affecting every bank, indicating that the attacks targeted specific vulnerabilities present in individual institutions rather than a universal flaw.
What should you do?
Security teams should monitor for unusual outbound connections to known AI service endpoints or residential proxies, which the attackers used to mask their origin. Open directories on attacker infrastructure exposed files and logs, suggesting that poor configuration can aid defense efforts. Organizations should review access logs for repeated connections from internal servers to external testing infrastructure.
Background: Generative AI
Generative AI examples include text completion, image synthesis and code generation. These systems predict the next likely token based on vast training datasets. Understanding the underlying mechanics helps you identify risks like data leakage and hallucination in your infrastructure.
Read the full guide: Generative AI Examples: Mechanisms, Risks and Control
What to do and how to stay safe: TAG-217
- Audit external-facing systems for open directories that might expose logs, scripts, or configuration files to public view.
- Monitor network traffic for connections to residential proxies that could indicate attempts to hide the origin of scanning activity.
- Review access logs for repeated internal-to-external server connections that may signal preparatory testing by threat actors.
- Once the vendor provides an update, apply patches to any systems found to have vulnerabilities exploited during the campaign.
Step-by-step guide: Social Engineering Warning Signs and How to Respond
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Which AI tools were used in the South Korean bank attacks?
The attackers used Claude Code, DeepSeek, Grok, and Zhipu AI alongside the ARTEX open-source penetration testing framework.
How many people were affected by the data breaches?
Public reporting indicates that personal information belonging to at least 68,000 individuals across seven financial institutions was exposed.
Is the threat actor TAG-217 confirmed to be state-sponsored?
No, Recorded Future described the actor as a suspected Chinese speaker but did not confirm any state backing.
