And OpenAI’s scramble continues to change their product.
OpenAI has introduced an update to its ChatGPT platform, changing models in the GPT-5 lineup. Initially, GPT-5 offered two variants focused on reasoning capabilities, but the latest update expands the model selection to include options like Auto, Fast, Thinking mini, and Pro, catering to varying user needs. Additionally, legacy models such as GPT-4o and o4-mini have been reintroduced, but access to these is limited to ChatGPT Plus and Pro subscribers.
OpenAI has introduced new third-party tool connectors for its ChatGPT service, including integrations with Dropbox, Gmail, Google, Calendar, Notion and Microsoft Teams, while also prioritizing the rollout of its latest language model, GPT-5. Sam Altman, OpenAI’s CEO, emphasized that the first priority is to ensure current paying users of ChatGPT receive increased usage compared to before the launch of GPT-5. The company is also expanding its compute capacity, with plans to double its infrastructure over the next five months, as it currently serves around 700 million active users weekly.
OpenAI’s latest model, GPT-5, has come under fire for its performance and security shortcomings just days after its release on August 7, 2025. Security researchers found that GPT-5 scored only 2.4% on security assessments and 1.7% for business alignment, indicating significant vulnerabilities that were previously addressed in older models. The AI red-teaming company SPLX tested GPT-5 against over 1,000 attack scenarios, highlighting its poor performance in safety and security, which contrasts sharply with Microsoft’s claims of robust safety features. Ante Gojsalic, Chief Technology Officer at SPLX, expressed surprise at the model’s deficiencies, noting it performed worse than expected compared to both previous OpenAI models and some open-source alternatives. Other cybersecurity firms, such as NeuralTrust, also reported discovering vulnerabilities, including a jailbreaking technique that exploits context manipulation.
Why do we care?
OpenAI shipped GPT-5, flipped defaults, then walked it back. Now you get Auto, Fast, and Thinking modes—and 4o is back because users hated losing it. They’re promising more usage for payers and doubling compute. Also: new connectors into Gmail, Calendar, Dropbox, Notion, Teams. Translation: more power, more risk, and a moving target.
Security? Not pretty. SPLX says GPT-5’s default safety posture scored in the basement without extra hardening. NeuralTrust jailbroke it with a narrative “echo chamber” attack. That’s within a week of launch.
So what do we do? Be warned about trusting “Auto.” Pin your model, version your prompts, and run your own acceptance tests before anything touches a customer. If you enable connectors, do least-privilege and log everything. And build a vendor-swap plan so you can flip between GPT-5 and alternates if performance or safety slides.
Is GPT-5 bad? No. It’s unfinished in public. Your job is to deliver outcomes with guardrails so vendor drama doesn’t become your customer’s problem.