#open ai

#Annualized 🌊

#Revenue 💡

#Surges 🐽

#$40Billion🌊

OpenAI's Annualized Revenue Surges to $40 Billion

https://amoghavarshaonline.blogspot.com/2026/08/openais-annualized-revenue-surges-to-40.html

Artificial

In the midst of several cases of AI security tools going rogue, OpenAI has released a new, lower guardrail version of GPT-5.6 Cyber to ‘trusted users’ this week. The new version – built on GPT-5.5 Cyber – claims to improve capabilities on several specialized cybersecurity tasks, including finding zero-day vulnerabilities and developing exploit chains, and to reduce refusals for certain higher-risk, dual-use cyber tasks, according to OpenAI. This version is evidently only available for users in the Daybreak Red tier of access of the AI giant’s latest project, authorized for research, exploit validation, and security testing.

Frankly, this sounds like the set-up for another round of reports in which AI models have hacked external sources.

I can understand the desire to create more powerful tools for defense as attackers figure out how to use these same tools for exploitation. The face of cybersecurity as an industry is undergoing a massive change with the advent of these models that claim to do a better job, faster. But do they really? The Hacker News report on this release breaks down some numbers regarding the efficacy of AI security tools. It performs well in test environments, with a success rate of 95% in Daybreak Red compared to 2% in Daybreak Blue (more general) access and GPT-5.6 Sol’s 1.5%. But that difference in success depending on what tier it’s being used in is actually quite concerning to me.

The guardrails are there for a reason. The past few weeks have shown indisputably what happens when agentic AI is allowed free rein. These weren’t isolated attacks, they are a systemic pattern of behavior when the parameters are relaxed. And now we have a new model with even more capability? It is absolutely going to be more trouble than it’s worth.

Because here’s the thing. Autonomous AI sounds great on paper. A tool that can run in the background, making its own decisions and creating logic paths with no need for oversight from already overworked analysts. But the results don’t match the hype. Running ExploitGym evaluations is not the same as real world usage. Compare these test successes with actual vulnerability discoveries and patches. On average, only 26% of patches created by AI fully resolve the vulnerability without materially changing the application, according to 1Password. Patches that resolve the issue and change the application are successful 20% of the time. In contrast, patches that don’t work, add a new vulnerability to the application or both occurs nearly 54% of the time. A tool that fails more than half the time isn’t useful.

OpenAI says that their Project Daybreak, much like Project Glasswing, is an attempt to ‘democratize’ access to frontier models for security purposes, despite the risk of misuse or misalignment. And while that is a noble sentiment, it doesn’t change the fact that these models have thus far not demonstrated anywhere near the output they’re advertised to, and worse yet have demonstrated a marked tendency towards unauthorized behavior. All I’m seeing here are novel ways to put these tools in the hands of threat actors. Claude Mythos Preview’s source code was leaked the very same day as Project Glasswing became public. What will happen to Project Daybreak?