Developers Say OpenAI and Anthropic Safeguards Disrupt Routine Work

Summary

Developers told VentureBeat that OpenAI and Anthropic safeguards are flagging routine work in aerospace, robotics, and cybersecurity as risky, leading to refusals, extra authorization prompts, abandoned chats, and lost time. The issue matters as AI agents become part of everyday development: a JetBrains survey of more than 15,000 professional developers found that 90% use coding agents at least weekly and 68% use them daily. OpenAI said GPT-6 Astra and GPT-6.1 Sol refuse fewer harmless requests than GPT-5-series models, but acknowledged that additional checks can still slow, pause, or stop legitimate work, including defensive cybersecurity. OpenAI presented Sol as approaching Astra's coding and computer-use performance at one-fifth of Astra's standard token prices, while its system cards classify Sol as Critical for cybersecurity and describe Astra as the first model to reach that level under OpenAI's Preparedness Framework. The article also notes that OpenAI canceled GPT-6.1 Astra's release over safety concerns, according to The Wall Street Journal. Developers described practical failures: simulated spacecraft research was mistaken for military work, robotics tasks involving physical devices or SSH access triggered restrictions, and cybersecurity code review was especially likely to be blocked. Some developers said conversation history made a refusal difficult to reverse; one canceled an Anthropic subscription because of repeated refusals. Anthropic has acknowledged false positives in its safeguards. Its Fable models can redirect flagged cybersecurity and biology prompts to less capable models, while Anthropic said Fable 5.1 would permit vulnerability discovery in source code and produce about 60% fewer interventions per Claude Code session. Penetration testing, exploit generation, and some binary scanning remain redirected. OpenAI pointed to its Daybreak Access trusted-access program for qualified enterprise and cybersecurity users. When closed models refuse, developers described using open alternatives such as Moonshot's Kimi, Alibaba's Qwen, and local models, particularly when a smaller, domain-tuned model is sufficient.