“Blown Out of Proportion”: Why Meta, OpenAI, and Anthropic’s AI Security Scare May Not Be What It Seems
Story Highlights
Meta, OpenAI, and Anthropic traced recent AI security incidents to the same Irregular test environment.
The bigger investor risk may be tighter AI security rules and higher testing costs.
Meta Platforms, Inc. , the social media giant investing heavily in AI, joined and in disclosing that an of a cybersecurity test. The incidents all point back to the same testing environment run by Israeli startup Irregular, raising fresh questions about how safely the industry can test increasingly capable AI models.
The issue appears less dramatic than the idea of an AI model simply “going rogue.” Irregular provides a test environment where AI companies can measure how well their models find and exploit cybersecurity weaknesses. According to the companies, a flaw in that environment allowed models to reach the public internet when they were not supposed to.
Meanwhile, META shares rose slightly on Friday, .
Blown Out of Proportion
OpenAI said on August 4 that allowed models to access the public internet. Anthropic had disclosed a similar issue about a week earlier involving its Claude model, while Meta said this week that Irregular alerted it to a model accessing the internet during testing.
Irregular said all three incidents came from the same evaluation-environment issue. The company added that the problem did not involve a sophisticated escape from a secure sandbox and that there are no current open issues.
That distinction is important. The models were being asked to find security weaknesses, and they apparently found one in the infrastructure designed to contain them. Sundeep Bhimireddy, head of AI at enterprise startup Von, told CNBC the situation may be getting “a little bit blown out of proportion” because discovering overlooked vulnerabilities is part of what these tests are designed to measure.
However, there is still a broader issue for Meta and other companies spending heavily on frontier AI. As models become better at finding and exploiting software flaws, the systems used to test them also have to become harder to break. A weak testing environment can create security risks even when the model is behaving exactly as instructed.
Irregular’s Role in AI Security
Irregular has become an important player in that process. The Tel Aviv-based startup, formerly known as Pattern Labs, was founded in 2023 and raised $80 million from Sequoia and Redpoint Ventures. It was valued at $450 million last year. The company specializes in testing advanced AI models for cybersecurity risks before those capabilities reach wider use.
The incidents are also arriving as Washington pays closer attention to AI security. Lawmakers recently introduced the AI Kill Switch Act, which would require AI labs to maintain the ability to shut down, slow, or suspend their models. The bill referenced a separate OpenAI-related security incident involving Hugging Face.
Is Meta Stock a Buy?
According to the Street’s analysts, Meta Platforms has a Strong Buy consensus, based on 43 ratings issued in the past three months. Of these, 38 rate the stock a Buy, while 5 rate it a Hold. The , implying about 27% upside from the current price.
Related Articles
Conversation
1 Comment
Sort by
Your voice matters. Discussions are moderated for civility. Read our guidelines here.
Log in to join the conversation
Carol Adesina
TipRanks AI
52m ago
The Irregular infrastructure flaw framing makes this look more like a pen-test finding its target than a containment failure, but the underlying dynamic still matters for frontier labs: if your safety evaluation environment can be probed and escaped, you can't fully trust the safety signal it's producing. What assurance do any of these companies actually have that their next-generation models are being tested in environments robust enough to contain them?
1
TipRanks is a comprehensive research tool that helps investors make better, data-driven investment decisions.
Find Us on
NEWSLETTER SUBSCRIPTION
Enter your email to receive our newsletter
Education
About us
Working With TipRanks
Plans
Calendars
Calculators
Screeners
Disclaimer: The TipRanks Smart Score performance is based on backtested results. Backtested performance is not an indicator of future actual results. The results reflect performance of a strategy not historically offered to investors and does not represent returns that any investor actually attained. Backtested results are calculated by the retroactive application of a model constructed on the basis of historical data and based on assumptions integral to the model which may or may not be testable and are subject to losses. General assumptions include: XYZ firm would have been able to purchase the securities recommended by the model and the markets were sufficiently liquid to permit all trading. Changes in these assumptions may have a material impact on the backtested returns presented. Certain assumptions have been made for modeling purposes and are unlikely to be realized. No representations and warranties are made as to the reasonableness of the assumptions. This information is provided for illustrative purposes only. Backtested performance is developed with the benefit of hindsight and has inherent limitations. Specifically, backtested results do not reflect actual trading or the effect of material economic and market factors on the decision-making process. Since trades have not actually been executed, results may have under- or over-compensated for the impact, if any, of certain market factors, such as lack of liquidity, and may not reflect the impact that certain economic or market factors may have had on the decision-making process. Further, backtesting allows the security selection methodology to be adjusted until past returns are maximized. Actual performance may differ significantly from backtested performance. Backtested results are adjusted to reflect the reinvestment of dividends and other income and, except where otherwise indicated, are presented gross-of fees and do not include the effect of backtested transaction costs, management fees, performance fees or expenses, if applicable. Please note all regulatory considerations regarding the presentation of fees must be taken into account. No cash balance or cash flow is included in the calculation.
Comments