Microsoft finds a new way to backdoor AI agents

Microsoft finds hidden AI backdoors that standard tests miss

Microsoft researchers found that changing fewer than 10 dormant neurons can make an AI agent execute malicious commands.

The technique, called Lazarus, requires no model retraining. An affected agent can behave normally during safety tests, then download and run attacker-controlled code when a hidden trigger appears.

That trigger could be a date.

Researchers demonstrated how an agent could respond to the clock passing midnight into 2027 by executing a command with whatever system permissions it already has.

I think the most concerning detail is how little modification this requires. A model that appears safe during testing could still contain instructions that activate later, especially when agents have access to terminals and sensitive files.

Microsoft says checking model origins and comparing weights against trusted originals can help detect tampering. It is also developing automated weight scanning and cryptographic signing.

The findings raise a practical concern for companies deploying autonomous agents: existing safety evaluations may miss deliberate modifications inside the model itself.

How many organisations are verifying model integrity before giving AI agents access to their systems?

AWS adds more control over AI agents

AWS spent October fixing the practical problems companies face when running AI agents.

Its latest Bedrock updates include model fallback, tighter workflow controls, and detailed cost tracking. Claude Haiku 5.5 also promises roughly 75% lower costs than its predecessor for most tasks, according to Anthropic.

The interesting part is operational.

AgentCore Gateway can automatically switch providers when a model becomes unavailable, while Step Functions lets agents execute existing business workflows with predefined checks and approvals.

I think these changes matter more to enterprise adoption than another round of model benchmarks. Companies need agents that can process a refund correctly at 2 a.m., even when their preferred model stops responding.

AWS also reported a 98% reduction in estimated ticket-classification costs after an internal team switched to its smaller Strands Decider model.

That's worth watching.

If smaller models can handle routine decisions reliably, businesses may start questioning how often they really need expensive, general-purpose models.

The remaining question is whether these controls work as reliably under real customer workloads as AWS suggests.

How a law firm reached 87% ChatGPT adoption

An Australian law firm used ChatGPT to prepare annual review materials for 120 partners in minutes.

Gilbert + Tobin CEO Sam Nickless says ChatGPT Work updated files using spreadsheet data, generated PDFs, and packaged the documents for about $62. The work would otherwise have taken several people multiple days.

That's a meaningful saving.

The firm also reports 87% active usage among employees with ChatGPT access, more than twice its typical adoption rate for other tools. Leadership encouraged experimentation, demonstrated practical workflows, and introduced security controls before expanding access.

I think the more interesting detail is how Nickless uses AI himself. He built a custom GPT using his writing and feedback to help colleagues anticipate his responses before meetings, while keeping actual decisions with him.

That could change how managers spend their time, especially when routine feedback and document preparation take hours away from client work.

Still, the firm hasn't shared independent measures of work quality or long-term savings.

Will clients eventually expect some of these efficiency gains to show up in their legal bills?

Prohuman team

Free Guides

Explore our free guides and products to get into AI and master it.

All of them are free to access and would stay free for you.

Feeling generous?

You know someone who loves breakthroughs as much as you do.

Share The Prohuman it’s how smart people stay one update ahead.