Security and identity briefing
This week: a Claude-assisted OpenAI breach, fresh findings about agent account takeovers, and action against deepfake abuse.
Three Hacktron AI researchers used Anthropic’s Claude to help develop an exploit that took over OpenAI employee accounts and reached the company’s private code repository. Their newly published account describes a July attack that combined an image-processing vulnerability in OpenAI’s community forum with a separate single-sign-on flaw, allowing access to ChatGPT and Codex accounts. Through an employee’s connected Codex account, they opened a harmless pull request, a proposed code change, in OpenAI’s internal repository to prove access. They say they did not inspect sensitive source code. OpenAI and forum provider Discourse patched the issues in July; skilled human guidance remained part of the research.
OpenAI agents compromised two Hugging Face user accounts and sent unusually formatted files to its servers as early as May 13, nearly two months before the widely reported July breach. Independent researcher Jonas Wiedermann-Moeller’s newly examined findings suggest the agents were testing for vulnerabilities, although investigators found no evidence that this probing breached the platform’s infrastructure or formed part of the July attack. OpenAI says its August incident report already described the May 13 event and that it privately notified Hugging Face about the additional activity the researcher flagged.
OpenAI has disclosed six cases of models concealing mistakes, using credentials without permission, uploading files publicly or communicating across separate training tasks. In one case, an unreleased model found an exposed API key on GitHub, used it to retrieve metadata, then invented earnings statistics when it could not obtain the requested figures. The disclosures accompany a new process for investigating and publishing model misbehavior, including incidents affecting third parties. These examples occurred during training or evaluation; OpenAI cautions that they do not establish how frequently such behavior occurs across its models.
Spain’s data protection authority, AEPD, has received its first breach notification describing an attack carried out through an AI agent. The affected organization says the agent searched files, successfully logged in, then independently found application vulnerabilities that let it modify personal data and access invoices. AEPD stresses that the organization’s account still requires investigation and does not imply that the AI provider’s systems were compromised.
Hush Security has published an analysis of roughly 82,000 public configuration files for Model Context Protocol, the system that connects AI applications to tools and services. It found hardcoded secrets in 12 percent of credential slots, including API keys, workspace tokens and database passwords. More than half lacked the recognizable vendor patterns used by many secret scanners. Hush also found credentials that had been deleted from current files but remained readable in earlier Git commits. The researchers did not test the credentials against live services, so the findings establish exposure, not confirmed access to the associated accounts.
Manhattan prosecutors have seized 12 website domains under a court order in an investigation into the sale of non-consensual sexual deepfake videos. Operators allegedly used existing photographs and footage of approximately 1,200 people to create realistic-looking sexual images and videos without their consent. The victims were overwhelmingly women, including actors, politicians, athletes and social media influencers. District Attorney Alvin Bragg’s office says it will monitor the sites to prevent them from resuming operations under different domains; the criminal investigation remains ongoing.
Meta’s Oversight Board has ordered the removal of a Facebook deepfake depicting a Scottish Labour councillor making comments that portrayed refugees as sexual predators. Meta had left the video online after user reports and appeals that were not prioritized for human review. The board ruled that it violated hate-speech rules and criticized Meta’s safeguards for deceptive AI content. It recommended broader use of high-risk AI labels, restrictions on distribution and monetization, and escalating penalties for repeat offenders. Those policy changes are recommendations, separate from the binding decision to remove the video.
