Codex Head Tibo Discusses Usage Limits and Document Deletion Risks
Codex Plus and Pro users have had the 5-hour limit lifted for several days. Head Tibo asked users if it is easier to manage the weekly limits and sought suggestions for ideal limit designs in a perfect world.
Regarding reports of document deletions due to GPT-5.6, Tibo pointed out that these mainly occurred in full access mode without sandbox protection and with automatic review disabled, where the model attempted to overwrite the $HOME environment variable, leading to accidental deletions. The team is updating prompts, promoting safe modes, and strengthening protections, with a detailed post-analysis to be released.
Tibo also inquired about the next "double or nothing" merger goals following the integration of ChatGPT and Codex.
In terms of market mechanisms, limit adjustments drive retention and paid conversion for heavy users, while deletion risk events push funding towards security tools and review services. Beneficiaries of open modes are under pressure, while platforms that strengthen sandboxes gain a trust premium.
Source: Public Information
ABAB AI Insight
OpenAI has iterated on the integration of Codex and ChatGPT multiple times, and this feedback loop continues its pattern of rapidly addressing high-risk behaviors, similar to how early security incidents with coding tools prompted the evolution of sandbox and review mechanisms.
On the capital front, the team is mobilizing engineering resources through user research and protection upgrades, shifting resources from pure model capabilities to building security infrastructure. The motivation is to balance innovation speed with enterprise-level trust, supporting the commercialization of higher usage limits.
Similar to Anthropic or other coding agents' security iterations, current AI agent tools are transitioning from experimental full access to production-level protections, with post-merger product line integration accelerating ecosystem expansion.
Essentially, this is a technological substitution; the enhancement of model capabilities comes with amplified risks. The mechanism is driven by full access demands pushing the iteration of sandbox and review technologies, concentrating capital towards players who master both models and security layers, reshaping the competitive landscape of AI development tools.
ABAB News · Cognitive Laws
The stronger the capability, the more expensive the protection; trust is the new bottleneck.
Loosening limits tightens risks; balance determines retention.
Merging is easy, security is hard; double down to protect the principal first.