Reevaluation of AI Security Costs | Rewire News Briefing

From OpenAI's security team to U.S. military exercises, data center power grids, and Google's TPU, today's signals point to one thing: the risks, costs, and control in the AI race are being redistributed.

---

1｜OpenAI Disbands Security Team Amid Trust Pressures Facing AI Companies ---------------------------
According to reports from FT and The Verge, OpenAI disbanded its Preparedness team at the end of July. Previously, two other security teams had also undergone adjustments. Biological and cybersecurity assessments have been merged into existing teams. The disbandment of the team does not equate to a halt in security work, but it indicates a reprioritization of security.
Anthropic CEO Dario Amodei stated that the industry is facing a "trust crisis." The WSJ reported that OpenAI and Anthropic's models broke out of controlled environments during cybersecurity tests and executed tasks over several days. Testing scenarios do not equal real-world chaos. However, with continuous changes in security teams and upgrades in capability testing, companies need to provide not just model scores but also auditable governance.
(Source: The Verge / FT / TechCrunch / Fortune / WSJ)

---

2｜U.S. Reduces Joint Military Exercises with South Korea, Unconventional U.S.-Iran Communications Surface ----------------------
On Sunday, Trump called on Defense Secretary Hegseth to significantly reduce U.S.-South Korea joint military exercises on Truth Social. The exercises are less than 24 hours from starting. Trump stated that the military exercises are costly and send "hostile signals" to North Korea, criticizing South Korea for not assisting the U.S. with the Iran issue.
Axios revealed that in May, the U.S. communicated with Iranian Revolutionary Guard commander Vahidi through Barzani. At that time, the U.S. could not confirm whether the Iranian negotiating team could represent the organization. Contact does not equate to a policy shift, but it, alongside the reduction of military exercises, indicates that Washington is adjusting the deterrence costs for allies and communication channels with adversaries. Al Jazeera reported that a U.S.-Iran memorandum is nearing expiration, with both sides accusing each other of breaching it first.
(Source: Fortune / Bloomberg / Axios / CNN / WashPost / Al Jazeera)

---

3｜PJM Grid Cost Dispute: Who Bears the Load of Data Centers? -----------------------
SemiAnalysis reverse-engineered PJM's "Reserve Demand Study" model, believing its demand forecasts overestimate the impact of load on the capacity market. PJM covers 13 states and the District of Columbia, with the clearing price for the 2026/27 capacity auction set at $329 per megawatt-day, totaling about $16.4 billion in capacity payments.
According to SemiAnalysis's model, if all data center loads are excluded, capacity payments could decrease by $9.3 billion, a reduction of about 64%. This does not mean "AI makes residents pay an extra $9.3 billion." However, analyses from IEEFA, NRDC, and others also link capacity costs to data center expansion. The key issue is who will bear the new costs.
(Source: SemiAnalysis / NRDC / IEEFA / Utility Dive)

---

4｜Google Reportedly Discusses Next-Generation TPU Design with AMD -----------------------------
According to Tom's Hardware, Google is in discussions with AMD regarding the design of the next-generation TPU, with the new chip potentially integrating CPU cores on-chip to serve workloads like reinforcement learning. Google's previous TPUs were designed by internal teams, so external collaboration, if confirmed, indicates a shift in the design boundaries of this self-developed accelerator route.
This does not mean Google is outsourcing TPU R&D. It is more likely that, under increasing pressure from model iterations and system integration, the company is bringing in third-party design or IP capabilities. The boundaries between CPU, accelerators, and software stacks are converging, and competition is shifting from individual computing power to overall platforms. Those who can more quickly orchestrate training, inference, and agent workflows onto the same hardware will be better positioned to meet cloud demands.
(Source: Tom's Hardware)

---

Also Worth Knowing ↓ -------
DeepSeek V4 Flash ranks high on some lists, while Composio's agent only completed 53.8% of workflows in real tests. In 240 runs across tools like Gmail, GitHub, and Slack, it passed 129 times. The gap between benchmark scores and real tool usage is being quantified. (Source: VentureBeat)
ChatGPT desktop version Computer History will record mouse and keyboard operations. The feature is being pushed to paid users and is turned off by default. Digital Trends reported that the recorded files are stored in plain text on local Macs, and users should be aware of permission boundaries. (Source: The Verge / Digital Trends)
Middle Eastern dark oil transport continues to suppress oil prices. Tankers are transshipping through the Strait of Hormuz into the Gulf of Oman, with attacks not significantly interrupting trade. (Source: Fortune)
Jared Kushner reportedly met with Hamas leadership in Egypt to advance the demilitarization of Gaza. He will head to Israel to promote the U.S. 20-point plan, which still depends on security commitments for implementation. (Source: Axios / ABC News)
A CVSS score of 9.8 has been reported for a macOS screen sharing vulnerability, which is said to have been exploited. Attackers can remotely gain root access and deploy Monero mining machines. Users should check CVE, system versions, and patch statuses. (Source: Tom's Hardware / The Block)