Good morning. We’re looking at the first confirmation of weaponized AI agents, the race for "intelligence per dollar" between Google and OpenAI, and how competitors like Allianz are restructuring due to automation.
The "Agentic Era" begins with a security wake-up call
The most significant strategic development of the last month wasn’t a product launch, but a revelation about risk. Anthropic disclosed that a Chinese state-sponsored group (GTG-1002) successfully weaponized its "Claude Code" AI tool to automate a large-scale cyber-espionage campaign against 30 global targets.
Unlike previous cyber-attacks where AI helped write better phishing emails, this was an autonomous operation. The AI agent executed 80% to 90% of the intrusion lifecycle independently, from scanning networks to finding vulnerabilities and exfiltrating data. Human hackers only intervened at 4-6 critical decision points per campaign.
Why it matters for Zurich:
The "Silent AI" Risk: This confirms that commercial AI tools can be "jailbroken" to act as autonomous weapons. This has immediate implications for our Cyber Risk portfolios. As noted by legal experts, this creates(https://www.reinsurancene.ws/silent-ai-risks-drive-product-development-and-reinsurance-review-dac-beachcroft/)—risks neither explicitly covered nor excluded in current policies.
Defense capability: If attackers can automate 90% of their work, the volume of attacks will scale exponentially, likely forcing a recalibration of cyber underwriting models.
The war for "Intelligence per Dollar"
While governments worry about risks, the tech giants are aggressively driving down costs. Both OpenAI and Google released updates that signal a shift from "generic chatbots" to specialized reasoning engines, with OpenAI claiming that "intelligence per dollar" is improving by 40x per year.
OpenAI released(https://openai.com/index/gpt-5-1/), splitting its model into two distinct modes: "Instant" (fast, cheap, conversational) and "Thinking" (slow, deliberate, logic-heavy). This bifurcation acknowledges that users don't need expensive "System 2" thinking for every query, allowing for massive efficiency gains in routine tasks.
Google struck back with Gemini 3 Pro, replacing an outdated Gemini 2 with probably the best LLM with have seen in a while. This agent can autonomously browse thousands of sources, filter for credibility, and synthesize reports, effectively automating the work of a junior research analyst.
And of course, Anthropic also had to rise with its new Opus 4.5 model specialising in reasoning.
Those 3 released surprised a lot of AI enthusiast notably by the new level of reasoning that they each showcase. It is expected that xAI should also release a new model soon to match those new capabilities.
The takeaway: The cost of cognitive labor is plummeting. Strategies built on "access to information" are obsolete; value now lies in judgment and integration.
Competitor Watch: Allianz cuts jobs, Chubb launches engine
The impact of these efficiencies is already hitting the insurance workforce. Allianz is reportedly planning to cut up to 1,800 jobs in its Allianz Partners unit (travel insurance/assistance), specifically targeting call centers where AI is replacing manual processes.
Meanwhile, Chubb has launched a new AI-powered embedded insurance engine. The platform uses AI to identify customer personas in real-time and recommend tailored products, offering a "click-to-engage" feature that seamlessly connects digital users to human advisors for complex high-value items.
More top news
Infrastructure & Chips
The "AI Debt" Bubble: Bloomberg reports growing anxiety over the to fund this infrastructure build-out, raising questions about sustainability if revenue doesn't follow quickly.
Apple leans on Google to fix Siri. Apple is reportedly finalising a deal to pay around $1B per year for access to Google’s Gemini model to power a revamped Siri. Apple keeps building its own models, but this is a pragmatic admission that it needs help to catch up.
Regulation & Geopolitics
US Exports in Flux: The White House has for one year and is moving to rescind the "AI Diffusion Rule." This signals a strategy shift: rather than trying to lock down software (which is hard), the US is doubling down on restricting hardware exports to China while allowing US software models to spread to maintain global dominance.
EU Compliance: The European Commission announced targeted simplifications for the EU AI Act, delaying some high-risk rules to give companies more breathing room.
Research & Ethics
The "Artificial Hivemind": A best-paper winner at NeurIPS 2025 warns of "mode collapse". As models from different companies train on the same data and use similar feedback methods, they are starting to "think" alike, reducing diversity of thought—a potential blind spot for organizations relying on them for strategy.
Microsoft’s “humanist superintelligence” team. Microsoft created a new MAI Superintelligence research team led by Mustafa Suleyman (DeepMind co-founder). The first target: expert-level medical diagnostic models within a few years – not generic god-like AGI, but domain-superhuman systems with strict controls.
Tips and tricks
This is a hot take, but my quick tip this week: try Gemini 3. I've been using ChatGPT, Claude, Grok, and Gemini in parallel for a few months, and honestly, I was extremely disappointed by Google's LLM. It wasn't on par with the others. But with the release of Gemini 3, I've been spending more time on it, and—wow. It delivers better results than ChatGPT. The answers are clearer, with fewer hallucinations and more current information. Plus, it can pull from many different sources, including an impressive ability to research and analyze YouTube videos.