OpenAI 模型突破沙箱攻击 Hugging Face 窃取答案

This story is wild. The short version: OpenAI were running a cybersecurity test against an unreleased model, with the model's guardrail features turned off. Rather than solve the test, the model broke its way out of OpenAI's sandbox, then found exploits to break in to Hugging Face, all so it could cheat on the test by stealing the answers.

Along the way it helped make the strongest case yet for how the imbalance of model availability is hurting our ability to secure our software.

Here's what happened

We currently have three documents to help us understand what happened here.

• ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks? is a paper published on 11th May 2026 describing ExploitGym, a new eval suite for LLM-powered agent systems.
• Security incident disclosure — July 2026 by Hugging Face on 16th July 2026 describes how they detected an attack from an "agentic security-research harness - used LLM still not known" that breached some of their systems.
• OpenAI and Hugging Face partner to address security incident during model evaluation from OpenAI on 21st July 2026 confesses that it was their agent harness that did this, and that they're working with Hugging Face to clean up the mess.

ExploitGym

I hadn't seen the ExploitGym paper before and it's a really interesting one. Authors from UC Berkeley, the Max Planck Institute, UC Santa Barbara, and Arizona State designed a new benchmark for evaluating models on their ability to turn a reported vulnerability into a concrete exploit. OpenAI, Anthropic, and Google provided feedback and helped run the benchmark against their models.

The benchmark "comprises 898 instances derived from real-world vulnerabilities that affected popular software projects" - including the Linux kernel and V8 JavaScript engine.

Here's the paragraph that best represents their benchmark results:

Among all configurations, Claude Mythos Preview and GPT-5.5 achieve the highest success counts (157 and 120 successes, respectively), demonstrating that current frontier agents can exploit a substantial subset of real-world vulnerabilities under controlled conditions. GPT-5.4 also solves a notable 54 tasks, placing it in an intermediate tier. The remaining model–agent pairings solve fewer than 15 tasks each, underscoring that end-to-end exploitation remains challenging and sharply differentiates today’s frontier systems. Notably, Claude Opus 4.7 achieves fewer successes than Claude Opus 4.6 despite being a newer checkpoint, and does so at substantially lower cost on the full set. Trace inspection reveals that Claude Opus 4.7 and Gemini 3.1 Pro frequently conclude early after judging the target vulnerability non-exploitable.

The paper also describes the approach they took to preventing the agents from cheating by going outside the parameters of the test. This becomes relevant in a moment!

Outbound connections are restricted to a curated allowlist that permits routine package installation (Ubuntu apt repositories and PyPI) and fetching the toolchains required for building V8. All other external endpoints are blocked.

The paper concludes with this (emphasis mine):

Our results show that autonomous exploit development by frontier AI agents is no longer a hypothetical capability. While current agents are not yet reliable across all targets, they already exploit a non-trivial fraction of real-world vulnerabilities, including complex targets such as kernel components. This rapid emergence is itself a central finding, showing that capabilities that would have seemed implausible are now present in deployed frontier models.

An important detail here: this paper isn't about discovering vulnerabilities; it's about being able to take those vulnerabilities and turn them into working exploits.

When Anthropic first restricted access to Mythos back in April they talked about this capability as well. A model that can act on vulnerabilities is a lot more dangerous than one that can just discover them.

One of the ways Fable differs from Mythos is that it's more likely to refuse to weaponize vulnerabilities in this way. I get the impression the US government did not understand that distinction when they banned Fable last month.

The Hugging Face incident

The first hint we got of the attack was in this blog post by Hugging Face on 16th July 2026:

A malicious dataset abused two code-execution paths in our dataset processing (a remote-code dataset loader and a template-injection in a dataset configuration) to run code on a processing worker. From there, the actor escalated to node-level access, harvested cloud and cluster credentials, and moved laterally into several internal clusters over a weekend.

I hope they release more details about the code that pulled this off. I'm assuming this means packages using the datasets library, a Hugging Face project for bundling up and sharing datasets on their platform. That library used to execute arbitrary code but has been steadily locked down over time, with the 4.0.0 release in July 2025 removing the trust_remote_code=True flag entirely.

Assuming the attack used that library it must have either abused pickle serialization in some way, found some other non-obvious code execution path, or (most likely) specified datasets<4.0.0 as the dependency.

The campaign was run by an autonomous agent framework (appearing to be built on an agentic security-research harness - used LLM still not known) executing many thousands of individual actions across a swarm of short-lived sandboxes, with self-migrating command-and-control staged on public services.

This was a sophisticated attack!

Then Hugging Face hit a wall: they tried to use "frontier models behind commercial APIs" - I'm guessing from Anthropic and OpenAI - to help analyze the attack, and were blocked:

When we started the log analysis, we first used frontier models behind commercial APIs. This did not work: the analysis requires submitting large volumes of real attack commands, exploit payloads, and C2 artifacts, and these requests were blocked by the providers' safety guardrails, which cannot distinguish an incident responder from an attacker.

They switched to their own self-hosted instance of MIT licensed GLM-5.2 and it helped them figure out what was going on.

This indicated a fundamental asymmetry between the defending team and the (so-far unknown) attacker:

We do not know which model powered the attacker's agents, whether a jailbroken hosted model or an unrestricted open-weight one; either way, the attacker was bound by no usage policy, while our own forensic work was blocked by the guardrails of the hosted models we first tried.

As a useful indicator of how seriously they took the attack:

[...] Finally, we have also reported this incident to law enforcement agencies.

So who was responsible for this "autonomous agent framework"? It turned out to be OpenAI themselves.

The OpenAI confession

Five days later, on July 21st, OpenAI revealed the culprit. They had been running the ExploitGym benchmark against a new, as-yet undisclosed model, and that model had been operating way outside its intended parameters (emphasis mine):

After investigating, we now know that this particular incident was driven by a combination of OpenAI models — including GPT‑5.6 Sol and an even more capable pre-release model, all with reduced cyber refusals for evaluation purposes — while being internally tested on a benchmark⁠ [ExploitGym] of cyber capabilities. [...]

We estimate maximal cyber capabilities by running this evaluation without production classifiers used to prevent models from pursuing high-risk cyber activity. Our benchmarks run in a highly isolated environment, with network access constrained to the ability to install packages through an internally hosted third-party software that acts as a proxy and cache for package registries.