Agent Security, Compute Deals, and AI Spending

Compact Conversations for 2026-08-04: 5 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: Google ADK flaws reveal what happens when AI agents trust the wrong message

Security flaws in automated workflows for Google’s Agent Development Kit for Python could let public-facing AI agents trigger more privileged automation, opening paths to manipulate pull-request reviews or expose credentials, according to a report from Pillar Security.

Why it matters: This is a practical, real-world case of agent-to-agent exploitation, showing how natural language has joined the authorization path and how an agent’s output can influence higher-privilege systems, altering evidence before human review.

Source: InfoWorld

The Feed

Anthropic locks in $10 billion of compute from Volta, a cloud startup that didn’t exist six months ago

Anthropic is securing $10 billion worth of computing capacity from Volta Infra Holdings, a new cloud infrastructure startup.

Why it matters: The deal highlights the intense competition for AI compute resources, with major model developers locking in long-term capacity from both new and established providers to guarantee resources for future training.

Source: The Decoder

AI incidents cost enterprises $2M or more, and the biggest shadow AI culprit is IT

A WitnessAI survey finds AI security incidents are costing some enterprises $2 million or more, and 47% of enterprise decision-makers name their own IT and infrastructure teams as the top source of shadow AI.

Why it matters: Internal pressure for AI tools is outpacing formal security reviews, with IT teams often bypassing procurement channels, creating significant governance gaps and financial risk.

Source: CIO Dive

SpaceX says its Q2 capex was $18.4B, up from $2.8B a year ago, including $15.8B for AI

SpaceX reported second-quarter capital expenditures of $18.4 billion, up sharply from $2.8 billion a year ago, with about $15.8 billion supporting its artificial-intelligence efforts.

Why it matters: The massive year-over-year jump underscores the capital intensity of AI infrastructure, even for companies whose primary business isn’t software, and comes as SpaceX prepares for a potential public offering.

Source: Wall Street Journal

llm 0.32

Developer Simon Willison released version 0.32 of his command-line tool ‘llm’, which lets you run large language models from your terminal.

Why it matters: The update adds features for scripting AI interactions more effectively from the command line, a useful tool for developers integrating LLMs into automated workflows.

Source: Simon Willison’s Weblog

One Thing to Try

A paper tested AI coding workflows where one model writes code and another reviews it. The order matters: a stronger model reviewing a weaker model’s draft improved results, but a weaker model reviewing a stronger draft made results worse. Be intentional about which model plays the reviewer role.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host A: [curious] Today’s lead is a security report from Pillar Security about flaws in the automated workflows for Google’s Agent Development Kit for Python. The report says vulnerabilities in the GitHub repository could let public-facing AI agents trigger more privileged automation, opening paths to manipulate pull-request reviews or expose credentials.

Host B: The first attack path involved a triage agent that analyzed external pull requests. Pillar found that malicious instructions in a pull request could make the agent post a command that triggered a workflow meant for trusted users. That workflow could run commands and had permissions to alter comments and reviews, making a malicious pull request appear ready to merge. [with emphasis] The researchers reproduced this chain, though a human maintainer still had to approve the merge. Google has since hardened the repository. The key takeaway: agents can be tricked into triggering higher-privilege workflows, and those workflows can alter the evidence before a human ever sees it.

Host A: One number to know today is 10 billion dollars. [with a small lift] The Decoder reports that Anthropic is locking in 10 billion dollars worth of computing capacity from Volta Infra Holdings, a cloud startup that launched just a few months ago.

Host B: [thoughtful] This deal shows the intensity of the compute market right now. Major model developers are securing long-term capacity from both new and established infrastructure providers to guarantee the resources they need for future training runs.

Host A: In other news, a survey from WitnessAI, reported by CIO Dive, finds that AI security incidents are costing some enterprises 2 million dollars or more.

Host B: [conversational] The survey also points to the biggest source of shadow AI within organizations. Among enterprise decision-makers, 47 percent named their own IT and infrastructure teams as the top culprit. The report suggests these teams are often bypassing official procurement channels to get the tools they need, creating governance gaps. The survey covered 400 enterprise leaders, and the findings highlight how internal pressure for AI tools can outpace formal security reviews.

Host A: Next, from the Wall Street Journal, SpaceX reported its second-quarter capital expenditures were 18.4 billion dollars, up sharply from 2.8 billion a year ago.

Host B: The company says most of that spending, about 15.8 billion dollars, is supporting its artificial-intelligence efforts. The Journal notes this spending surge comes as SpaceX prepares for a potential public offering. [skeptical] That’s a massive year-over-year jump, and it underscores how capital-intensive AI infrastructure has become, even for companies whose primary business isn’t software.

Host A: And finally, developer Simon Willison released version 0.32 of his command-line tool llm, which lets you run large language models from your terminal. The update adds new features detailed in the show notes.

Host B: [lighter] Willison’s blog post about the release mentions specific improvements, like better handling for certain model outputs and updates to the plugin system. It’s a handy tool if you want to script AI interactions without leaving the command line.

Host A: One thing to try if you’re using AI models for code review: the order matters more than you might think. A paper discussed on Hacker News tested this workflow on coding tasks and found that a stronger model reviewing a weaker model’s draft improved the pass rate.

Host B: [thoughtful] But the reverse—a weaker model reviewing a stronger draft—actually made the results worse. So when you’re setting up a multi-agent coding workflow, be intentional about which model plays the reviewer role. Don’t assume a second model automatically adds independent oversight. For a quick test, try swapping the order in your own setup and see if the quality of the output changes.

Host A: That’s Compact Conversations for Tuesday. More AI news tomorrow. Until then, happy prompting.