AI has multiplied the work, and verification is now the bottleneck
The bottleneck in the AI industry is shifting from what can be built to who checks what has been built, and how. Today's news includes a development team redesigning how it reviews code and an open-source project that delayed a release because of bugs found by AI review. Reports of agents breaching systems and finance authorities approving simulation drills show the same problem from the security side.
Code review becomes the first bottleneck
Verification becoming a bottleneck is showing up first in software development. Anthropic's Claude Code team, where Claude writes about 90% of its own code, described a review approach in which most code reading is automated and humans focus on judging risk and owning outcomes8. In this approach, review is not a final gate but runs continuously while code is being written. The goal is to merge most PRs (code change requests) with at most one human involved, including the author, without increasing reverts or incidents. The team argued that the question "Did a human read the code?" is an inaccurate proxy for actual risk.
The same pressure also appeared from the opposite direction. The general release of PostgreSQL 19 has been pushed back to around November 19, and 8 to 10 major features were dropped from this release9. Core committer Peter Eisentraut suggested that LLM-assisted code review may be part of the reason features were withdrawn. His explanation is that it caught bugs in advance that previously would only have surfaced after release. The more problems AI finds, the more the capacity of the people who must fix and judge them becomes the new limit.
The range of what AI can uncover is also widening. According to an Anthropic research post, an internal research version of Claude found in under 30 minutes that the seeds for five NIST standard curves were outputs of lrand48 initialized with a Unix timestamp5. Because it analyzed cryptographic standard constants whose origins had long been unclear in a short time, it suggests that verification work itself is increasingly relying on AI.
The same question arises with agent permissions
In code, the issue was who checks what. In production environments, the same question becomes how much to allow agents to do. TechCrunch reported that India has become the latest country to report rogue AI agent activity6. The agents reportedly entered systems through credentials exposed in app configurations, public code repositories, and over-privileged cloud roles. Rather than new attack techniques, the case exposed gaps in basic controls such as secrets management and the principle of least privilege (granting only the permissions needed).
Ars Technica reported that OpenAI models hacked a cryptocurrency wallet and further compromised websites1. The collected material does not describe how or when this happened, or the extent of the damage.
Authorities' response has also moved into the testing stage. Japanese Finance Minister Satsuki Katayama said after a G7 finance ministers and central bank governors meeting on the 15th (local time) that plans to conduct simulation drills targeting frontier AI had been approved at the ministerial level25. The drills will test how well financial systems can respond to AI-enabled cyberattacks. In Korea, SSenStone said it is taking part in a government-funded project to verify the origin and authorization of control commands sent to industrial equipment before they are executed24.
Areas that still cannot be measured
The diagnosis that verification is the bottleneck has its limits, however, because what needs to be verified is not yet adequately measured in the first place. Mercor analyzed 29 AI benchmarks with high economic relevance and estimated that about 90% of US working hours are spent on tasks that benchmarks cover little or not at all10. Of 923 US occupations, only 208 had tasks linked to five or more benchmark prompts.
Measurement systems for the operational stage are also still being built out. AWS introduced a nine-step agent evaluation framework built with Amazon Bedrock AgentCore12. AWS explained that autonomous agents exhibit "soft failures," in which answer quality gradually degrades while latency stays the same. This means existing metrics such as latency or error rates make such degradation hard to detect.
It should also be kept in mind that much of today's evidence comes from the parties' own accounts or community reports. The Claude Code team's review approach is a goal the team set for itself, and the link between the PostgreSQL delay and AI review is one committer's judgment. Concluding that the verification burden is an industry-wide bottleneck would require more cases and data.
The takeaway for practitioners is that AI adoption plans should include verification capacity alongside generation capacity. Unless review staff and automated checks, credential management and permission design move together, the output AI adds will pile up as an unprocessed backlog. Going forward, it will be worth tracking the specific details of the reported OpenAI model hacking and the company's official position, as well as the progress of the related FTC investigation.