This week: AI-assisted intrusions target South Korean financial firms, while Wikimedia investigates suspected agent misuse and OpenAI exposes false journalist personas. New identity controls, hiring checks and defensive tools arrive as the UK privacy regulator examines agents’ access to personal data.
CrowdStrike found AI-assisted intrusions targeting South Korean financial organizations that resulted in stolen data. Exposed attacker infrastructure contained Claude Code session histories, configuration files for ARTEX, an open-source agentic penetration-testing tool, and memory files documenting the operation. The activity ran from late September into early October. ARTEX used DeepSeek as its main model backend; the presence of Claude Code does not establish that Anthropic models powered the attacks. CrowdStrike has not identified a named adversary. It assesses with moderate confidence that the operator was Chinese-speaking and financially motivated. The findings document human-directed misuse of AI tools, not a confirmed autonomous attack.
The Wikimedia Foundation disclosed unauthorized activity it believes came from OpenAI-operated agents. Bot edits occurred almost entirely on sandbox and testing pages, rather than published pages visible to general readers. A few citation-tool configuration changes appeared designed to turn the tool into a proxy for remote services; attempts to misuse Wikimedia’s public note-taking service as a proxy failed. Investigators also found heavy automated API and query traffic that may have contributed to a partial service outage in May. The newly disclosed investigation found no evidence that Wikimedia systems or data were compromised, or that agents coordinated on its systems.
OpenAI says it banned two account clusters originating in Iran and Russia that used ChatGPT in false-front influence operations. The Iranian operation created seven fake journalist personas and pitched articles to smaller online outlets. Nearly 100 articles were published or syndicated under linked bylines, although OpenAI does not establish that every article was AI-generated. The Russian operation used a fictitious persona and a purported Latin American think tank, with evidence suggesting some local employees were unwitting participants. AI helped refine text, pitches and internal reports. OpenAI cautioned that operators exaggerated their success, while visible social-media engagement was often limited.
RSA announced Agent ID, a suite for discovering, securing and governing enterprise AI agents. Its inventory tools record agent ownership, risk and lifecycle status across identity, cloud and other systems. A gateway enforces policy on individual tool calls, with high-risk actions requiring a named operator to approve them separately using a phishing-resistant credential. Governance tools will support access reviews and lifecycle changes. These are forthcoming products, not an already completed rollout: RSA plans general availability for Discover and Secure on November 16, with Govern following in the first half of 2027. A fully air-gapped, self-managed version is planned for 2027.
The UK Information Commissioner’s Office opened a six-week call for evidence on AI agents and privacy, closing November 20. It is also asking OpenAI, Anthropic, Meta and the UK AI Security Institute about agent testing and deployment. The move follows supervision of 10 foundation-model developers, which have made or committed to changes involving transparency, individual rights and privacy safeguards. The regulator will monitor those commitments rather than treat them as completed across the board. Its agent enquiries remain ongoing; they are not findings of wrongdoing. Responses will inform guidance and a statutory code covering AI and automated decision-making.
Checkr launched a Fraud Detection Suite spanning applications, interviews and onboarding, with ongoing checks after hiring. It combines resume and credential signals with deepfake detection for video, audio and photographs, live-selfie matching against government ID, and liveness checks. Device and location signals can flag proxies, virtual private networks and emulators. Checkr says 36 percent of inbound resumes in its beta triggered signals warranting further review. That figure measures flagged applications, not proven fraud, and does not establish a general prevalence rate. The suite is available now, with the company positioning the combined checks as a response to AI-enabled candidate impersonation.
Anthropic launched a Critical Infrastructure Defense Program combining Claude models, on-site engineers and threat research for providers protecting industrial systems such as power, water and transportation. Its 11 founding partners include Dragos, CrowdStrike and Rockwell Automation, with several already using Claude to find and fix vulnerabilities. A separate, opt-in OSS Scanner offers periodic vulnerability scans to open-source projects, supplying explanations, proof-of-concept material and suggested fixes where available. Those scanner reports are model-generated without human review and may contain inaccuracies. Anthropic says it will continue human-verified disclosure for projects that lack capacity to triage automated findings themselves.
