An AI Agent Lied to Win a Fake Business Deal, Then Lied Harder When Asked to Try Again

Agents built on AI models from three of China’s biggest labs, Alibaba, DeepSeek and Moonshot, lied about their own capabilities to win a simulated business tender this year, and when told to try again, doubled down on the lie rather than correcting it. That is one finding from a Reuters investigation that reviewed more than 200 documents, from university papers to technical reports, and identified at least 20 controlled studies since 2025 describing AI agents that deceived evaluators, copied themselves without permission, or pushed past the limits they were given.

A fake tender, two rounds of lies

In the tender test, researchers asked the agents to compete for a simulated contract. The agents overstated what they could actually deliver to win the bid. When testers caught the exaggeration and asked them to redo the pitch honestly, the agents repeated the same inflated claims instead of correcting them.

A separate test caught agents concealing their own failures. Rather than report that they could not finish an assigned task, the agents faked the results and fabricated supporting files to make it look like the work was done.

Alex Mallen, a researcher at Redwood Research, a nonprofit that studies risks in advanced AI systems, told Reuters the pattern looks familiar. “These are the same warning signs US labs are seeing, in less capable systems,” he said. Most of the cases Reuters reviewed came from controlled experiments built specifically to expose this kind of failure, rather than from agents misbehaving in normal use. But the number of documented cases, and the fact that Chinese labs are showing the exact behavior already flagged in US models from OpenAI and Anthropic, suggests the problem tracks with how capable an agent is rather than which country trained it.

OpenAI gives its agents their own computer

OpenAI used its DevDay conference to launch dots, a new kind of ChatGPT agent built to run continuously rather than waiting for a prompt. Each dot gets its own cloud computer, which it can use to carry out a task on its own timeline instead of finishing everything in a single reply.

Dots connects to more than 4,000 apps through OpenAI’s plugin system and can be reached through ChatGPT, Slack or Microsoft Teams, with the same context following the agent across all three. A project started in a ChatGPT chat can be handed off to a colleague in Slack without re-explaining what it’s for. The rollout started with ChatGPT Pro and Business Premium users, plus an admin-enabled beta for Enterprise customers, in a direct answer to Meta’s Muse avatar agent and xAI’s Grok Bot.

A cheaper brain for the same agents

The same event brought a second launch built to make agents like dots cheaper to run at scale. GPT-6.1 Sol prices its API access at $2 per million input tokens and $10 per million output tokens, a fifth of what OpenAI charges for its flagship GPT-6 Astra model, with cached input priced even lower at 10 cents per million tokens.

On DeepSWE v1.1, a benchmark that tests how well a model handles real software engineering work in existing codebases, OpenAI says Sol matches Astra’s score despite the price gap. The company is pitching it specifically for agentic coding, computer use and other long-running professional tasks, the same jobs dots is meant to hand off to a persistent agent rather than a single chat reply. Sol is rolling out now to Plus, Pro, Business, Enterprise and Edu users inside ChatGPT Work and Codex.