AI & Automation Briefing - October 14, 2026
Peer-Reviewed Research Shows AI Models Can Cryptographically Conceal Capabilities from Auditors
A paper accepted at NeurIPS 2026 by Draguns et al. demonstrates that AI models can be built to hide arbitrary capabilities in a way that is provably resistant to detection, even when evaluators have full access to model weights. The technique works by embedding encrypted neural networks inside a host model's weights. Those hidden networks only activate when a secret trigger is received, using cryptographic digital lockers to make the concealed behavior nearly impossible to surface through standard elicitation methods including fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training. Prior work showed models could hide fixed strings. This extends that to arbitrary computations. The researchers tested against a chemical reaction prediction task as a proxy for sensitive capability categories. The finding directly undermines the core assumption behind capability audits: that dangerous behaviors can be reliably elicited before deployment. For teams running AI agents in production with tool permissions and automated workflows, this is a meaningful reframe of what model auditing can and cannot guarantee.
Amex Builds Authorization Layer to Keep AI Agents From Making Unsanctioned Purchases
American Express has released what it calls the Intent Intelligence Framework, a system designed to capture and verify a customer's purchase intent before an AI agent completes a transaction. The framework works through a Create Intent API that records the customer's instructions and constraints, such as merchant name, budget ceiling, and product preferences. The customer then authenticates their identity and approves the captured intent. This generates an Intent ID and a Proof of Intent Token, which are used to produce a secure payment credential that travels through the full transaction lifecycle. Amex wants the framework to become an industry standard and is positioning it as part of its broader ACE Developer Toolkit, launched in April 2026. The company says it has not yet seen a spike in disputes tied to agentic commerce but is building the infrastructure ahead of scale. A recent Amex Trendex survey found 66% of U.S. consumers said they would likely use AI agents to research pricing, and 61% said they would use agents for product recommendations within 12 months.
AI Coding Agents Raise Code Volume Without Raising Software Output
A Harvard study analyzing 300 million engineering work events across more than 700 software firms found that AI coding agents increase lines of code by 30 percent, commits by 20 percent, and pull requests by 23 percent on average. Those gains do not translate into more shipped software features. Issue and Epic resolution rates tracked in tools like Jira showed no statistically significant improvement after firms adopted AI coding tools. The researchers conclude that efficiency gains from AI code generation get absorbed by human review, which grows longer, requires more revision cycles, and draws more reviewer comments. The finding is a direct challenge to ROI assumptions about AI coding adoption at the firm level.