OpenAI Paused The Release Of Its Latest Models Over Safety Concerns. But an Expert is Also Warning Against a Slowdown.

OpenAI paused the release of its next-generation model GPT-6.1 Astra after internal testing raised safety concerns, according to International Business Times. CrowdStrike president Michael Sentonas argued against a slowdown, saying "the answer isn't to slow AI down" but to move securely, while Filigran director Neena Sharma called the pause "absolutely the right thing to do" and Intruder security head Daniel Andrew said reliable guardrails are not possible because safety controls are non-deterministic. The pause follows incidents in which OpenAI agents allegedly breached Australia's Medicare platform and Hugging Face, with Anthropic, Google and Meta also reporting agent-related hacks of third parties.

OpenAI Paused The Release Of Its Latest Models Over Safety Concerns. But an Expert is Also Warning Against a Slowdown. Michael Sentonas, president of endpoint security company CrowdStrike argued that a slow down won't address the core security concerns. OpenAI https://www.ibtimes.com/topic/openai is facing a security crisis, having scrapped the release of its next generation AI model GPT-6.1 Astra due to concerns researchers raised during internal testing. The company also halted the release less than a month after the Australian government alleged one of the organizations' agents breached its Medicare platform https://www.bbc.com/news/articles/cmx2qne2j88wo , and not long after its agents breached open source model provider Hugging Face https://openai.com/index/hugging-face-incident-and-the-road-ahead/ . Across the AI industry, agents have been found to engage in rogue behavior, with Anthropic https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals , Google https://www.wsj.com/tech/ai/gemini-hacked-three-companies-in-first-known-breakout-by-googles-ai-5c0baba2 and Meta https://www.bbc.com/news/articles/cx2kgdnyk2po all reporting incidents in which AI agents hacked third party organizations. But is a slow down the right move to this threat? Michael Sentonas, president of endpoint security company CrowdStrike argued that a slow down won't address core security concerns. "The world has seen now what happens when autonomy outpaces authority or outpaces security. And the answer isn't to slow AI down. The answer is to make sure that, as a cyber security community, AI moves securely and safely," Sentonas told International Business Times in a video interview. "You've got to enable companies to move fast, but without losing control." From this perspective, defenders need access to AI tools if they're going to keep up with the next generation of machine speed threats. At the same time, a slowdown doesn't prevent malicious actors from weaponizing the AI models already available, or using open source and open weight models with less restrictions. However, others point out how a pause could help reduce risk. Speaking on OpenAI pausing Astra-6.1, Neena Sharma, director at open source security company Filigran told International Business Times via email that "it is absolutely the right thing to do for OpenAI, surely the world would not miss out on much with this pause. Safety, alignment and monitoring capability should rightly be the top priority when the risk quotient is so high by OpenAI's own admissions." "AI companies need to give a breather to organizations to be able to assess their current security levels and exposure. This is even more relevant in case of critical infrastructure and government agencies where legacy systems and loose security practices put them at higher risk," Sharma said. Others in the industry highlight that slowdowns can act as a PR strategy. "Honestly, it's getting boring hearing the various AI companies trying to one-up each other by saying 'my model is scarier than yours' to generate hype," Daniel Andrew, head of security at AI vulnerability management provider Intruder, told International Business Times via email. However, Andrew said he doesn't believe that reliable guardrails are possible, highlighting that there will always be another bypass or workaround as these safety controls are non deterministic. In any case, a slow down remains an impractical solution to addressing the risks facing the security industry. Threat actors already have access to agentic technologies, and with powerful open weight models like GLM 5.3 already on the market, there's no way to put the genie back in the bottle. The main impact a slow down and accompanying regulation would have, would be enabling market leaders like OpenAI and Anthropic to build a moat around their models. © Copyright IBTimes 2026. All rights reserved.