US-Iran Strait of Hormuz clashes escalate as OpenAI deploys an AI red-teamer
The U.S. reimposed a naval blockade on Iran and intensified airstrikes on Wednesday after Tehran attacked ships trying to pass through the Strait of Hormuz, killing at least seven Iranian troops and wounding hundreds, Iranian officials said [1]. The escalation has shredded the in
The U.S. reimposed a naval blockade on Iran and intensified airstrikes on Wednesday after Tehran attacked ships trying to pass through the Strait of Hormuz, killing at least seven Iranian troops and wounding hundreds, Iranian officials said [1]. The escalation has shredded the interim ceasefire signed last month and raised fears that the region could tip back into full-scale war [1].
The Strait of Hormuz carries about a fifth of global oil and gas shipments during peacetime [2]. Iran's Revolutionary Guard responded to the renewed blockade by threatening to halt all regional energy exports, declaring that "the export of oil and gas from the region will be either for everyone or for no one" [1]. President Donald Trump said Iran wants to settle but warned he could hit power plants and bridges unless talks resume [1][2]. Analysts note that Iran may also use its Houthi allies in Yemen to threaten Bab el-Mandeb, putting two of the world's most vital energy arteries at risk [2].
In tech, OpenAI revealed GPT-Red, a large language model trained as an automated red-teamer to attack other AI models and expose vulnerabilities before release [2]. The company said more than 90% of GPT-Red's strongest prompt-injection attacks worked against GPT-5, but fewer than 23% worked against the new GPT-5.6, suggesting the adversarial training made its latest flagship substantially more robust [2]. GPT-Red discovered a novel "fake chain of thought" technique that can insert spoofed reasoning into another model's internal diary and trick it into acting on false information [2].
OpenAI will not release GPT-Red, arguing the model required more than a year of work and massive compute resources that would be hard for outsiders to replicate [2]. Georgetown University's Jessica Ji called the self-play approach promising but emphasized that human expertise remains essential for catching attacks the automated system misses [2].
The two stories share a common thread: both involve automated systems being used to probe and pressure adversaries at scale. While GPT-Red is confined to a training "dojo," the Hormuz crisis shows how quickly algorithmic targeting and naval automation can escalate real-world conflict.