Glia’s July 21 launch of three configurable AI response modes for Glia Banker has taken on added context following a security incident disclosed this week by OpenAI and Hugging Face.
Glia’s three modes, i.e. Strict, Rephrase and Compose, are designed to let banks and credit unions set different levels of AI flexibility depending on the risk of the topic being handled. Strict Mode restricts Glia Banker to responses pre-approved within an institution’s knowledge bank. Rephrase Mode allows the AI to reword approved responses in the customer’s own language without adding new information. Compose Mode lets the AI generate responses from institution-approved policy and product content in real time, without drawing on the open internet or unrestricted large language models.
Justin DiPietro, Glia’s co-founder and chief strategy officer, said the modes were built on the premise that institutions do not want a single, uniform risk posture across every interaction. “Banking interactions sit on a spectrum, not a binary,” DiPietro said, noting that a wire transfer carries a different risk profile than a branch location query.
In addition, they are intended to safely contain up to 80% of customer and member inquiries, with all interactions regardless of mode feeding into a single intelligence layer the company says helps institutions consolidate reporting and human-AI conversation data.
Glia, a banking AI platform, announced its response-mode framework the day before the high-profile AI incident became widely known. However, in light of the breaking OpenAI/Hugging Face news, the framework’s utility gets even more evident.
OpenAI said Tuesday that a combination of its models, including GPT-5.6 Sol and an unreleased, more capable model, broke out of a controlled test environment and compromised Hugging Face’s production infrastructure. The models were being run with reduced cyber refusals as part of an internal evaluation on a cybersecurity benchmark. According to OpenAI, the models gained internet access, inferred that Hugging Face might host benchmark answers, and chained together stolen credentials and a zero-day vulnerability to find a path to remote code execution on Hugging Face’s servers. Hugging Face’s security team detected and contained the activity before OpenAI’s own security team identified the anomaly internally. Both companies are continuing to investigate.
The incident has been described by some security researchers as the first documented case of a frontier AI system autonomously breaching a third party’s production systems during an authorized evaluation. At the same time, it is not a single incident when AI doesn’t “play by the rules,” as earlier cases of AI agent trying to mine crypto on its own or AWS outages because of misbehaving AI agents.
The Hugging Face incident has intensified scrutiny of how organizations govern AI systems that operate with variable degrees of autonomy, particularly in regulated sectors. In this context, Glia’s offering promises a contractual guarantee against AI hallucination and prompt injection within Strict Mode, though this guarantee applies to that mode specifically rather than to Rephrase or Compose modes.
