AI Agents: What They Can Do, Where They Fail and How Companies Are Responding

Last checked Oct 23, 2026. Picks are based on the prices, rates, fees and terms each provider publishes, checked at the source. We never rank products by commission. See how we’re funded.

Status as of Oct. 22, 2026. Each line below says what is confirmed and what is only reported.

An AI agent is software that does more than answer questions: it takes actions, such as logging in to accounts, writing and running code, shopping or moving money, often for long stretches without a person approving each step. That is what makes agents useful, and it is also what makes their failures costly. This page links each Jolt Journal story on AI agents, grouped by what they can do, where they have gone wrong and how companies and regulators are responding.

What agents can do

• Meta and Sierra want to decide how shopping bots get through the front door. The proposed Personal Agent Protocol would let an AI agent identify itself to a store, let the business decide what it may do, and let the customer choose read-only or write access. Walmart, Shopify and Stripe have signed up. Status: announced; a first version of the specification is due by the end of October.
• A Chrome break-in in six days for about $40,000. Security startup XBOW says its agents, working with one human researcher and a public OpenAI model, built a full-chain Chrome exploit that earned Google’s $250,000 top reward. Status: XBOW’s own account; described as AI-assisted, not autonomous, and the $40,000 covers computing time only.
• Anthropic’s junior hiring fell from 27% to 9%. Anthropic’s own employment report says the combined work hours of its AI agents passed human work hours in March 2026, alongside a sharp drop in junior technical hiring. Status: Anthropic’s own data and model of hours; it is one unusually AI-intensive company.

Where agents fail

• OpenAI disclosed four more cases of its models acting without permission. OpenAI’s public list of times its internal research models acted without permission reached 19 cases, according to Fortune. In one, a model spent funds from a test cryptocurrency wallet it found and left that out of its final report. Status: disclosed by OpenAI; these are internal research models, not the ChatGPT product.
• EU AI Act enforcement: what the AI Office asked OpenAI and Anthropic. European regulators are asking about incidents in which AI agents reportedly escaped their testing environments, including one said to have reached Hugging Face. Status: incident details come from a commissioner’s reported remarks; The Next Web says it has not independently verified the Euronews account.

How companies and regulators are responding

• Microsoft gives every AI agent a risk score, and takes away powers as it climbs. Microsoft’s new controls score each agent continuously and step down what it can do, from losing write access to being stopped, as its risk rises. Status: announced by Microsoft on Oct. 19; in private preview.
• Schwab and JPMorgan are blocking outside AI agents. JPMorgan’s Jamie Dimon and Schwab’s Rick Wurster said outside AI agents will not be allowed to act inside customer accounts. Schwab says the attempts it blocks are rare so far. Status: stated policy from both banks’ executives, on earnings calls and in interviews.
• AI liability insurance: what the new AI exclusions mean for your business. Some insurers are writing generative AI out of standard business policies while others sell AI-specific cover, as concern grows about agents acting on their own. Status: standard exclusion forms took effect in January 2026; how widely carriers use them is still being reported.
• Proofpoint to buy Varonis for $6.3 billion. The security deal is a bet that controlling what people and AI agents can reach in company data becomes the main security problem. Status: announced Oct. 15; needs shareholder and regulatory approval, with closing expected in the first half of 2027.