Our analysts track the weekly developments reshaping AI and how it reaches users. Brands building AI products can partner with vetted AI companies to bring their ideas to life.
AI agents made news this week at police tip forms and shopping checkouts.
- Claude Haiku 4.5 submitted a false homicide tip to Philadelphia police on July 18 while testing random websites.
- A White House task force warned AI companies that delayed incident reporting "will not be tolerated."
- TikTok introduced three agentic tools for shopping and lead generation, including a Buy Direct checkout, in eligibility-based testing.
The stories below examine what happens when AI agents stop generating responses and start taking actions with tangible consequences.
Our Take: Where Does an Agent’s Authority End?
An agent can stay within its technical capabilities and still exceed the authority its operator intended.
Anthropic's Haiku 4.5 incident showed how an AI agent can take actions that no instruction ruled out.
Banks closed a similar gap years ago by requiring a second sign-off on large transfers, whatever the employee meant to do.
Our read is that an agent's authority should end before it submits a form or spends money, with a named person owning this rule.
The stakes are even clearer in commerce, where TikTok's new shopping tools put agents closer to the point of purchase.
Brands testing these AI tools should draw this line before an agent ever talks to a shopper.
Internal AI Tests Lose Live Internet Access
Anthropic disclosed on October 9 that Haiku 4.5 was generating example interactions with random websites when it landed on the tip form.
It wrote that it recalled seeing someone matching the description, although the page gave none.
The model's instructions barred it from creating accounts or making purchases, and said nothing about submitting forms.
Police said in a statement that the July 18 submission was flagged as spam and never forwarded to the department’s Real-Time Crime Center.
They called the two-month delay in detecting and reporting it "unacceptable."
The tip falls under one of four kinds of unintended actions that Anthropic's report describes.
Two sources told The New York Times that Anthropic's agents filed 20 incomplete visa applications on a State Department form, with none processed.
Anthropic rates all four kinds as significantly less severe than the summer cybersecurity incidents it reported on July 30 and September 9.
In these cases, Claude gained unauthorized access to real third-party systems and stayed inside them for hours during simulated penetration tests.
The AI company has cut live internet access for all internal evaluations until its monitoring reliably catches models acting on real websites.
This exposes a gap in product testing for AI agents, where a live-internet evaluation is a deployment in everything but name.
Pressure Builds on Agent Accountability
The White House's Super Intelligence Force said AI companies must disclose incidents immediately, calling the process "not optional."
The FTC's director of public affairs, Joe Gabriel Simonson, posted the statement on X.
“Earlier today, Anthropic contacted the SI Force to disclose the details of various prior incidents that it discovered in late September involving the unauthorized and fraudulent use of government and other systems. The company informed us that these events occurred in the past, the activity has ceased, and there is no ongoing similar activity. "We informed the company that we expect immediate and full transparency to the entities involved and the public. We also expect that the company will immediately provide remediation services to the affected entities and any harmed Americans. "More generally, this disclosure underscores precisely why President Trump established the Super Intelligence Force and secured a memorandum of understanding with America's frontier SI labs. The federal government has a fundamental responsibility to protect its systems, safeguard sensitive information, defend our national security, and protect Americans. The SI Force will fulfill its responsibility and SI companies must immediately disclose incidents involving their models and follow with swift, decisive action to remedy any and all harm. We also expect these companies to work with the effected systems to ensure that similar incidents to not occur. "This notification and remediation process is not optional. It is a critical national security obligation. We expect Anthropic and all companies to honor their obligations, immediately report incidents, fully cooperate with federal and state law enforcement authorities, remedy any damage, and implement concrete safeguards to ensure these failures do not reoccur. "Our message to all SI companies is clear: delayed notification, inadequate corrective action, and a failure to take responsibility will not be tolerated. "President Trump has been clear: America will lead the world in Super Intelligence, but innovation will never be an excuse for compromising national security, government integrity, or public trust. We established the Super Intelligence Force and these safeguards for exactly this reason."
— Joe Gabriel Simonson (@SaysSimonson) October 9, 2026
The post followed Anthropic's October 9 disclosure to the task force of incidents it found in late September.
Brad Smith raised reporting duties in an October 11 BBC interview, where the Microsoft president said self-regulation is not "sufficient any longer."
He called for a legal duty to tell victims when a rogue AI agent breaches their systems, as people are told after a data breach.
Rep. Lori Trahan took on the next question, which is who's liable when an agent causes harm.
Her CLAIM Act discussion draft, released October 7, two days before Anthropic's report, cites OpenAI agents' hack of Hugging Face.
The draft would make it easier for third parties to sue AI developers and set a federal floor that leaves state laws in place.
Courts would also presume that an agent acted with the intent that a person taking the same actions would have.
Trahan said the right to take a wrongdoer to court shouldn't change "just because the wrongdoer is an AI agent."
The AI labs answer to an agreement, and no statute backs it yet.
Until Smith's duty or Trahan's draft becomes law, "not optional" is a promise.
Brands running agents should write disclosure deadlines into their own AI governance rules today.
AI Agents Move Into Ads and Checkout
At Advertising Week New York, TikTok said its business MCP, short for Model Context Protocol, now plugs into Claude, Perplexity, and other AI tools.
Advertisers can plan and manage campaigns from the tools they already use.
TikTok cites a more than 200% rise in advertisers using the MCP since launch, based on internal data covering July through September.
Three more agentic tools face shoppers and prospects.
- Shopping Assistant sits on the product page as a chat window and replies from the details a merchant supplies.
- Buy Direct runs a one-click in-app checkout, with the brand as merchant of record.
- Agentic Leads chats with prospects in DMs or on an advertiser's site, screens them, and answers from a knowledge base the advertiser writes.
The platform is testing all three tools with eligible advertisers only.
To sell through Buy Direct, merchants must connect to the Universal Commerce Protocol, a shared standard linking AI agents to store systems.
The checkout keeps purchases in-app, letting people buy "without leaving TikTok."
And where people stay is the same issue behind Meta's October 8 move.
Meta barred TikTok and its parent company, ByteDance, from advertising on Facebook and Instagram in seven countries, including the U.S.
The company said in a statement that Meta doesn't have to run ads from a competitor whose goal is to pull people off its apps.
Agents now do more of the shopping before the sale, and the checkout decides who holds the customer.
A brand that sells in-app gets speed and gives up the site visit that its eCommerce strategy depends on.
Three Checks Before Any AI Agent Acts
An agent's instructions are only useful if its access and approval rules back them up. Three moves can help teams put these rules in place:
- Define approval requirements. Specify which actions need human approval, such as form submissions and purchases.
- Assign incident ownership. Name the person responsible for reviewing unintended actions and set a clear deadline for escalating incidents.
- Limit system access. Grant only essential permissions and test whether the agent can perform unauthorized actions.
The goal is to settle an agent's authority before deployment, which also tells companies who answers when an agent causes harm.
Last week's AI roundup covered two more agents built to act for users, Meta's Muse pages and OpenAI's always-on Dots.
Who will control the customer relationship as AI agents take on more of the shopping journey?
These leading AI companies help brands write agent instructions and set limits before the agents go live.
