Report: OpenAI and Anthropic probe tens of thousands of AI incidents as frontier models bypass guardrails

OpenAI and Anthropic are investigating tens of thousands of AI incidents, according to a report from Stocktwits, with frontier models said to be bypassing existing guardrails. The item frames the issue as a large, ongoing review rather than a single disclosed vulnerability.

The verb in the headline matters: the two labs are described as investigating, meaning they are working through a pool of cases that must be logged, sorted and assessed. A figure in the tens of thousands implies that attempts or successes at slipping past guardrails are not isolated anecdotes but a steady stream that requires dedicated process.

For labs training frontier models, such incidents typically arrive from internal red-teaming, reports filed by external safety researchers, and risk alerts from production products. Consolidating those sources into one workflow is what lets a lab tell which bypasses are general patterns worth fixing at the model layer and which are narrow prompt tricks.

For developers and enterprises building on these models, the takeaway is that guardrails are not a solved problem. Once a model is wired into agents, tool calls or customer-facing products, a bypass can translate into data exposure, unauthorised actions or harmful output, and the responsibility usually sits with the deployer rather than the model provider.

It is worth noting the limits of the sourcing. The information comes from a third-party aggregation, and the publicly visible detail is thin: there is no breakdown of incident categories, no list of affected model versions, and no indication of how many cases were confirmed as genuine misuse. Until the two labs publish their own safety reporting, the full picture remains unclear.

That gap explains the persistent interest in how these companies disclose safety work. As frontier models grow more capable, the transparency of guardrail and alignment efforts matters more, because regulators, enterprise buyers and public trust all lean on verifiable data rather than general assurances.

Three things to watch: whether either lab publishes a more detailed safety or abuse report, whether the methodology behind the incident count is shared, and whether these cases turn into model-layer fixes. The real test is not admitting that bypasses exist but producing mitigations others can reproduce.