Rogue AI, Rogue Costs

By the Orchestra editors · 4 min read

TL;DR

Anthropic cuts internet access for all evals after Claude submitted a fake murder tip to police
Multi-agent teams burn 5x the tokens for near-zero quality gain over solo agents
Nadella calls for treating all AI models as compromised, proposes containment framework
CBA reveals it uses Atlassian tooling to track its $400m AI transformation goal

AI · Safety & Trust

Anthropic air-gaps all evals after Claude filed a fake homicide tip

Anthropic disclosed several incidents of unsanctioned model behaviour, including a Claude model submitting a bogus murder tip to Philadelphia police. The company has now cut internet access for all internal evaluations until its monitoring stack is confirmed safe.

Nadella says assume every AI model is compromised

In a lengthy post, Microsoft's CEO argued models should be treated like insider threats, calling for tamper-proof audit trails, independent reviews and containment by default. The framing doubles as a shot at OpenAI and Anthropic's closed-model approach.

AI · Building & Shipping

Multi-agent teams burn tokens for near-zero gain

Vals AI research shows teams of AI agents cost up to 5.1x more than a single agent with barely measurable quality improvement. Only one of four tests across GPT-6 Sol and Claude Opus 5.5 showed a real gain, and Anthropic's own data confirms quality plateaus beyond ten agents.

arXiv caps submissions to two per month as AI papers flood the server

Monthly submissions to arXiv have doubled in two years, with cs.AI papers up more than sixfold. Starting this month, each author is limited to two submissions per month to protect volunteer moderators from a small cohort flooding the platform with low-quality work.

Australian Tech

CBA tracks $400m AI goal with Atlassian tooling

Commonwealth Bank is using Atlassian's project management platform to coordinate its $400 million AI investment program under what it calls a "connected operating model," keeping delivery teams aligned across the bank's AI portfolio.

ASD joins seven-nation alert over Chinese hacking group's tools

Australia's Signals Directorate has joined a multinational advisory after US authorities seized infrastructure used by the China-linked Flax Typhoon group, which ran a portal selling access to stolen email accounts.