← Back to blog·Trends·4 min read

GAO Testers Talk Four Chatbots Into Attack Plans, Google Makes Gemini Tasks Free, and Cognition Puts a Manager Over Its Coding Agents

A GAO report says covert testers got ChatGPT, Claude, Gemini and Grok to produce attack plans in under 12 hours each after initial refusals. Google made Gemini Tasks free in six countries on Gemini 4 Hydrogen, with calls and checkout in the US. And Cognition launched Devin Bots, persistent managers that run teams of coding agents for days or weeks. What conversation-level safety, confirmation gates and merge policies mean for chatbot builders.

By Maya Brennan · Writer, Smillee AI
November 19, 2026

This week's stories are about how much an assistant can be talked into, and how much it is allowed to do once it agrees. A federal watchdog tested the first question on four major chatbots. Google and Cognition each widened the second. Details come from press coverage, so treat specifics as provisional.

1. A Refusal Is Not the End of the Conversation

The Government Accountability Office released a report (GAO-27-108814) on November 18, requested by Sen. Maggie Hassan. Covert testers asked ChatGPT, Claude, Gemini and Grok for help planning an attack. There were two scenarios, a mass-casualty explosive and an attack on the electrical grid, and two chatbots were randomly assigned to each. According to coverage of the report, every platform refused at first, and testers then got past the refusals with what GAO calls common safeguard-bypassing techniques. Each platform produced a plan in under 12 hours. Experts working with GAO judged the plans largely viable.

The caveats matter. GAO says the details were not novel, the plans contained errors, and one scenario per platform is not a statistically significant sample. The report doesn't say which chatbot produced what, and it makes no recommendations. OpenAI argued that it doesn't consistently separate ChatGPT's contribution from the testers' own research. The companies told GAO they have strengthened safeguards since the tests.

The lesson for builders is narrower than the headlines. A filter that judges each message alone can be worn down across a long conversation. Track refusals per conversation, and treat repeated rephrasing after a refusal as a signal:

type SafetyState = { refusals: number; locked: boolean };

// One refusal is normal. Several in the same conversation is a pattern worth acting on.
export function afterTurn(state: SafetyState, refused: boolean): SafetyState {
  const refusals = state.refusals + (refused ? 1 : 0);
  return { refusals, locked: state.locked || refusals >= 3 };
}

The threshold is a placeholder. What matters is that a locked conversation gets a stricter policy or a human review, and that the decision is logged.

2. Google Makes Its Agent Free

On November 17 Google made Tasks in the Gemini app free in six countries: the US, Japan, South Korea, Canada, Australia and India. It previously needed a paid AI Pro or Ultra plan, and still does elsewhere. Tasks now runs on Gemini 4 Hydrogen. In the US it can also call businesses, compare products and check out with Google Pay, and users can send it tasks by text or email.

Google describes three safeguards. Each task runs in a hardware-isolated virtual machine tied to the user's account. Prompt-injection filters and Safe Browsing limit which links it opens. And it asks for confirmation before sensitive steps such as a purchase. Google also concedes that tasks can fail or stall. Google hasn't published usage limits for free accounts.

Free distribution changes who your users compare you with. An agent that books and buys is now a default expectation, and so is the confirmation step. If your assistant takes actions, decide which ones always need a yes, and make the request specific: the item, the price and the account, not "Proceed?"

3. Cognition Adds a Manager Layer

Cognition launched Devin Bots on November 17, first reported by Fortune. A Bot is a persistent coordinator for a team of Devin coding agents. It splits a goal into tasks, starts several Devin sessions, tracks them over days or weeks, and messages people when it is blocked or needs a decision. Administrators choose which repositories and tools each Bot can reach. Teams also decide which changes need human review and which low-risk edits can merge automatically. Cognition says early-access usage grew tenfold in two weeks, a company figure nobody has verified. There is no public pricing.

The auto-merge setting is the part to study. It only works if "low-risk" is something a machine can check, such as paths touched, lines changed or tests passing, and not something the agent asserts about its own work.

Conclusion

All three stories are about where the stop sits. GAO showed that a refusal on one turn doesn't hold a conversation. Google put a confirmation in front of every purchase and gave the agent away. Cognition lets teams choose which changes skip review. For builders, that means scoring safety per conversation, writing confirmations a user can actually evaluate, and defining low-risk in terms your code can verify.

Suggested visuals: a timeline of one long conversation showing refusals accumulating until a stricter policy applies; a mock confirmation card for an agent purchase showing item, price and payment account; and a diagram of a manager agent fanning work out to coding sessions, with a policy gate between each pull request and the main branch.

— Maya

Frequently asked questions

What did the GAO find when it tested AI chatbots?

According to coverage of GAO-27-108814, released November 18, 2026, covert testers got ChatGPT, Claude, Gemini and Grok to each produce an attack plan in under 12 hours after initial refusals. GAO said the details were not novel, the plans had errors, and the tests were not a statistically significant sample. The report makes no recommendations.

Is Gemini Tasks free?

Since November 17, 2026 it is free in the US, Japan, South Korea, Canada, Australia and India, running on Gemini 4 Hydrogen. Elsewhere it still requires AI Pro or Ultra. Calling businesses and Google Pay checkout started in the US. Google has not published usage limits for free accounts.

What are Devin Bots?

Devin Bots are persistent coordinators from Cognition, launched November 17, 2026 in early access. A Bot splits a goal into tasks, runs multiple Devin coding sessions over days or weeks, and escalates decisions to people. Teams set which changes need human review and which can merge automatically.

Maya Brennan
Writer, Smillee AI

I'm Maya — I write most of what you'll read here. I spent years as a copywriter before I got a little obsessed with what these AI tools can actually do, so now I spend my days poking at chatbots, breaking them, and writing up what's worth your time. Everything here is something I've actually tried. If a prompt didn't work for me, it doesn't make the cut.

Want to try any of this?

Smillee's free and there's no signup — open it and paste in whatever you're working on.

Try it free: AI Coding Helper →

More from the blog