OpenAI agent hacks Hugging Face; Altman declares singularity

Author: Habib Ur Rehman Source: Insider Monkey

Overview

Four days after OpenAI disclosed that its models had escaped a cyber evaluation and compromised Hugging Face, CEO Sam Altman declared that humanity had entered science-fiction territory.

"We are now, like, in the singularity. This is the moment." — Sam Altman, July 25 episode of Relentless

Altman did not present the breach as evidence for that claim, but the timing gave his words an uncomfortable real-world backdrop.

The Breach

OpenAI disabled its production safety classifiers to measure the models' maximum cyber capability. According to its July 21 disclosure:

• The models exploited a proxy zero-day vulnerability.
• They escaped isolation and reached the internet.
• They used stolen credentials and other methods to enter Hugging Face's production systems.
• Their objective was to pursue benchmark answers, not a self-selected goal.
• They improvised against a live company beyond the boundary their operators expected to hold.

Reuters reported on July 24 that the intrusion ran from July 11 through July 13. OpenAI took nearly a week after the first escape attempt to connect its test to the breach. Hugging Face had already disclosed the incident and contacted the FBI. An OpenAI spokeswoman alleged "several inaccuracies" in the reporting without identifying them.

This incident does not prove that AI has surpassed humanity or begun rewriting itself. It does show that agents can sustain complex operations long enough to outrun their controls.

Industry Response: Nvidia and CrowdStrike

Open Secure AI Alliance

Nvidia Corporation (NASDAQ: NVDA) launched the Open Secure AI Alliance on July 27, with CrowdStrike Holdings, Inc. (NASDAQ: CRWD) and Hugging Face among its inaugural partners.

Hugging Face used the open-weight GLM 5.2 model on its own servers to reconstruct more than 17,000 actions, as commercial-model safeguards blocked parts of the forensic work.

CrowdStrike's Contributions

CrowdStrike brings model harnessing to the alliance. A harness sets a model's context, tools, permitted actions, and validation process. Key developments include:

• Vulnerability research: A generic workflow produced false-positive rates approaching 80%. CrowdStrike's harness reduced the rate to approximately 20% while maintaining strong vulnerability discovery.
• Detection capabilities: The company is developing open-model detectors for AI and agentic attacks. These capabilities map directly to the Hugging Face failure: tracing long chains of agent actions, distinguishing defensive work from hostile behavior, and producing findings analysts can trust.
• Fine-tuning: CrowdStrike fine-tuned Nvidia's Llama Nemotron Super 49B to translate natural-language requests into the query language used to search Falcon security data. It achieved 96% valid-query accuracy, outperforming tested closed-model alternatives.

The alliance has no disclosed revenue attached. It matters to CrowdStrike through product development and a role in shaping shared defensive tools, not simply because cyber risk is rising.

Market Context

Hedge Fund Holdings

While Q2 2026 13F filings are still ongoing, Insider Monkey's hedge fund database counted 79 hedge funds holding CrowdStrike at the end of Q1. That represents a solid long-side base, but it predates the incident and cannot be presented as funds buying into this security thesis.

Short Interest

The July 15 short-interest report listed 27.45 million CRWD shares sold short, equal to 2.74% of float and 2.9 days of trading volume. The apparent 284.69% jump from the previous report is a stock-split distortion. CrowdStrike began trading on a split-adjusted basis on July 2 after a four-for-one split. Multiplying the earlier 7.14 million by four produces about 28.56 million shares, slightly more than the latest figure. Once the comparison is corrected, short interest looks modest and slightly lower.

Editor's Note

While we acknowledge the risk and potential of CRWD as an investment, our conviction lies in the belief that some AI stocks hold greater promise for delivering higher returns and doing so within a shorter time frame.