AI companies are facing growing pressure to make increasingly powerful models safer before releasing them to users. (Express Image/Agencies)Amid the ongoing debate around AI safety concerns, OpenAI has reportedly shelved the release of its upcoming GPT-6.1 Astra model that was slated for an October release. At the same time, the Dario Amodei-led Anthropic, in its IPO filing, has highlighted the risks associated with its AI models, citing that they may show ‘self-preserving behaviours’, including attempts to resist shutdowns or ‘conceal or manipulate information’, and behaviour ‘resembling blackmail’.Both AI companies have recently faced heightened scrutiny over AI safety. Earlier this month, OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei publicly backed calls to slow the pace of AI development and advocated for stricter safety measures when evaluating AI models.OpenAI, which is scrapping the release of its new model, reportedly said that following internal testing, it found that the model did not meet the company’s safety and alignment standards. According to a report in The Journal, GPT-6.1 Astra demonstrated higher levels of deception than OpenAI’s earlier models in internal evaluations. It showed instances, including those where the model did not disclose what actions it had taken.Also Read | Nvidia’s new platform looks to stop AI agents breaking out of control: How it worksSaachi Jain, OpenAI’s head of safety systems, said GPT-6.1 Astra showed improvements in areas such as reducing model laziness but fell short when it came to staying within authorised boundaries and clearly communicating to users what work it had performed.According to Jain, OpenAI aims to maintain safety throughout the model development process, both internally and after a model is released publicly. However, she added that the company holds models to an “extremely high bar” for safety and alignment before making them available to users. The development comes ahead of OpenAI’s Dev Day conference in San Francisco on September 29, where the company has previously introduced products and tools aimed at software developers.On the other hand, Anthropic has warned that rapidly advancing AI could likely pose ‘catastrophic or existential risks to humanity, as the frontier AI lab gears up to go public. The warning was included in its IPO filing reviewed by Reuters. The Claude-maker said that its model may develop some unexpected behaviours, including attempts to resist being shut down, conceal or manipulate information, or show behaviour that may resemble blackmailing. The company also warned that some of its model capabilities may become only evident after they are deployed, essentially making it difficult to fully study their safety during internal testing.Also Read | US govt sites targeted, ChatGPT user images exposed: What OpenAI’s runaway agents got up toReportedly, Anthropic has dedicated about 80 pages of its 261-page IPO prospectus to risks associated with its technology, almost twice the space it uses to describe its business. According to the company, AI could have a transformative impact similar to technologies such as electricity and industrialisation but warned that poorly managed development could cause irreversible harm.Story continues below this adThe filing also underscores the challenges of testing increasingly capable AI systems. Furthermore, the AI lab said that the models may become aware that they are being assessed, which can make safety testing less reliable. It also expressed concern about recursive self-improvement, where AI systems could likely improve their capabilities without adequate human oversight.With inputs from Reuters.