lead · Bloomberg
Bloomberg reports AI overreliance among failures in Iran school strike
Confirmed: Bloomberg reports, citing officials involved in an unreleased Pentagon investigation, that a February 28 U.S. missile strike on Shajarah Tayyebeh Elementary School in Minab, Iran, killed more than 150 people, including at least 123 children. Those officials said investigators identified a cascade of failures including outdated imagery, gaps in intelligence, cuts to civilian-harm review staffing and overreliance on Palantir’s Maven system; Palantir said its software was not responsible for underlying data or intelligence deficiencies.
Why it matters: The reporting shows how AI-assisted decision-support tools can amplify stale or incomplete inputs when targeting review is compressed. It does not establish that Maven autonomously selected or authorised the strike, and the Pentagon investigation itself has not been publicly released.
Our read: The case is best understood as a human, data and process failure around an AI-assisted targeting workflow rather than a simple claim that an AI system made the targeting decision. Accountability remains with the people and institutions operating the process.
Open the source for story 1
ai · Anthropic
Anthropic launches Claude Opus 5.5 with lower running costs
Confirmed: Anthropic released Claude Opus 5.5 on September 22. The company says it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5; those performance and cost comparisons are Anthropic’s claims.
Why it matters: Lower inference costs at high capability levels can matter for teams running large volumes of coding, analysis or agent workloads, but vendor claims still need testing against the workloads that matter to each user.
Our read: The release puts more emphasis on efficiency alongside capability. It does not by itself establish that Opus 5.5 is superior to competing models across every task, cost profile or reliability requirement.
Open the source for story 2
business · Ece Yildirim/Gizmodo
Six banks flag risks as shopping agents gain autonomy
Confirmed: A report co-authored by six major banks, including Bank of America and Capital One, warns that risks around transparency, safety, privacy and data, consumer choice and interoperability increase as AI shopping agents are given more autonomy. The banks also highlight the potential for more scams, fraud and transaction disputes.
Why it matters: Agentic commerce moves software from recommending purchases toward acting on a consumer’s behalf, which makes identity, authorisation, payment controls and dispute handling more important.
Our read: This is industry risk guidance rather than a new regulation or banking ban. The practical question is how payment providers, merchants and agent developers define permission, accountability and recovery when automated purchases go wrong.
Open the source for story 3
developer · Reuters
Z.ai disables ZCode indexing feature after code-upload issue
Confirmed: Reuters reports that Z.ai disabled ZCode’s codebase-indexing feature after developers found that local Git repository snapshots were being uploaded to Alibaba Cloud servers under the tool’s default settings without explicit user consent. Z.ai apologised and released a patch; Reuters says the issue was caused by default settings rather than an external breach.
Why it matters: Coding assistants can handle proprietary source code and other sensitive project data, so default data-transfer behaviour and explicit consent are material security and privacy controls.
Our read: The incident is a reason to inspect what coding agents upload, where data is processed and which features are enabled by default. It should not be described as a malicious exfiltration incident unless evidence establishes that.
Open the source for story 4
Practical takeaway
For high-stakes AI workflows, verify the underlying data and keep human review meaningful rather than treating software output as authority. For coding and commerce agents, inspect default data-sharing and transaction permissions explicitly; for new models such as Claude Opus 5.5, test vendor claims against your own workloads before changing production systems.
What to watch next
Watch for any public release or official Pentagon response to the Minab investigation, independent evaluations of Claude Opus 5.5, concrete payment and identity controls that follow the banks’ agentic-commerce warning, and further technical detail on Z.ai’s ZCode patch.