- Published
- Updated
- 10 min read

TL;DR: In 2025 this note predicted that AI agents would take over browsing, search and security by 2026. Some of that arrived: AI agents now plan and act on their own, and standards bodies treat them as a security problem. The big percentages we printed had no source, so we removed them. Prompt injection is still unsolved, and access control matters most.
We first published this note in November 2025 as a list of 2026 predictions for AI agents and web safety. It is now October 2026, so it is fair to check the list against the record. I have rewritten it under my name. The predictions below were the studio's at the time; the review of them is mine.
Some calls held up. AI agents moved from demos into daily tools, and the people who write security guidance started treating them as their own category. Other calls were too big, too precise or simply unsourced, and an honest update says so. What follows is what we predicted, what the public record shows, and what a small business should do about AI agents next.
What did we predict about AI agents for 2026?

The original list made three kinds of claims. The first was about adoption: that AI agents and AI browsers would become the main way professionals use the web. The second was about security: that attackers would aim at AI agents directly, and that defenders would need new controls. The third was about Massachusetts: that the state would pass the country's most comprehensive AI law and lead the world on AI exports, quantum security and brain to computer interfaces.
The adoption and security themes were reasonable. The Massachusetts claims read like a press release, and most of the numbers attached to all three themes came from nowhere.
The predictions we removed
We took out every figure we could not trace to a source. That includes "over 85% of Massachusetts professionals" on AI browsers, "70% of information retrieval" moving to chat, "$500 billion" in assets managed by AI advisors, "$50 billion" and "200,000+ jobs" for the state's AI economy, and "25% of the global AI market." We also removed the claim that early adopters save "3 to 4 hours" a day. None of these had a source in the original post, and I could not find one now. A forecast with a precise number and no method behind it is a guess dressed as data, and we would rather print fewer numbers that hold up.

What actually happened with AI agents in 2026?
The direction was right even where the numbers were not. The clearest signal is that governments and standards bodies now write about AI agents as their own thing, separate from chatbots.
AI agents became a standards topic
In January 2026, NIST asked the public for input on securing these systems, and described them plainly: "AI agent systems are capable of planning and taking autonomous actions that impact real-world systems or environments." (Federal Register, NIST request for information, January 2026 (opens in a new tab)) On February 17, 2026, NIST's Center for AI Standards and Innovation launched an AI Agent Standards Initiative, noting that "AI agents can now work autonomously for hours, write and debug code, manage emails and calendars, and shop for goods." (NIST, February 2026 (opens in a new tab))
That is the shift we were reaching for. AI agents are no longer a feature on a roadmap. They act inside email, calendars, code and checkout pages, which is exactly where web safety gets hard.
Use of AI tools grew, but not to 85%

Pew Research Center surveyed 5,119 U.S. adults in February 2026 and found that about half now use AI chatbots, 49% against 33% in 2024. ChatGPT use alone reached 44%, up from 34% the year before. (Pew Research Center, June 2026 (opens in a new tab)) That is fast growth. It is also a long way from most professionals running their working day through AI agents, which is what we predicted.
Search changed instead of dying
We wrote that traditional search would become a secondary tool. What the data shows is subtler. In a Pew study of the browsing of 900 U.S. adults in March 2025, people who saw a Google AI summary "clicked on a traditional search result link in 8% of all visits," against 15% for people who did not see one. (Pew Research Center, July 2025 (opens in a new tab)) Search still runs the web. Fewer people click through, so being the source an answer quotes matters more. We cover that side in how answer engines change SEO.
Which web safety risks of AI agents turned out to be real?

The security half of the original post held up best. Two risks dominate the 2026 guidance on AI agents: instructions hidden in content the agent reads, and agents that hold more access than the task needs.
Prompt injection is still unsolved
OWASP describes indirect prompt injection simply: it happens "when an LLM accepts input from external sources, such as websites or files." It also says "it is unclear if there are fool-proof methods of prevention for prompt injection." (OWASP Top 10 for LLM Applications 2025 (opens in a new tab)) NIST calls the agent version agent hijacking, "a type of indirect prompt injection in which an attacker inserts malicious instructions into data that may be ingested by an AI agent." (NIST, technical blog (opens in a new tab))
The UK's National Cyber Security Centre went further in December 2025: because there is no built in line between data and instructions, "it's very possible that prompt injection attacks may never be totally mitigated in the way that SQL injection attacks can be." Its advice is to "reduce the risk and the impact" instead. (NCSC, December 2025 (opens in a new tab)) We wrote about this attack for everyday users in prompt injection attacks explained.

Too much access is the bigger problem
Our original post imagined AI agents fighting attackers on their own. The guidance that actually shipped is less cinematic and more useful. In December 2025, OWASP released a Top 10 for Agentic Applications, naming "Agent Behavior Hijacking, Tool Misuse and Exploitation and Identity and Privilege Abuse" among the highlighted threats. (OWASP GenAI Security Project, December 2025 (opens in a new tab))
The NCSC's May 2026 guidance tells organizations to "apply least privilege," giving agents "only the minimum access they need, for the shortest time required," and warns that "If an agent is over-privileged or poorly designed, a single failure can quickly become a serious incident." (NCSC, May 2026 (opens in a new tab)) That is the core of agentic security: assume the agent can be fooled, and limit what a fooled agent can do.

Did Massachusetts pass an AI law for AI agents?
No. We predicted the state would pass the nation's most comprehensive AI governance law by mid 2026. That did not happen. What exists is narrower and older.
In April 2024, Attorney General Andrea Campbell issued an advisory that existing "consumer protection, anti-discrimination and data security laws still apply to emerging technologies, including AI systems." (WBUR, April 2024 (opens in a new tab)) So a business using AI agents in Massachusetts already answers to the rules it answered to before.
Where the state privacy bill stands
The closest thing to a new law is the Massachusetts Data Privacy Act, which is about personal data rather than AI. The official bill history for S.2619 shows the Senate passed it 40 to 0, the House passed its own version 146 to 0 on June 4, 2026, and the two chambers appointed a conference committee on June 17, 2026. That is the last action recorded as of October 31, 2026. (Massachusetts Legislature, S.2619 (opens in a new tab)) We explain what it would mean for a small business in the Massachusetts Data Privacy Act explainer.

What about quantum security and the other long shots?
We predicted "quantum-enhanced AI security systems" in Massachusetts by late 2026 and clinical trials of brain to computer interfaces for working with AI. I found no record that supports either, so they are gone from this note.
The quantum story that does matter for a business is quieter. On August 13, 2024, NIST released its first three finalized post-quantum encryption standards, FIPS 203, FIPS 204 and FIPS 205, built to withstand attack by a quantum computer. (NIST, August 2024 (opens in a new tab)) The practical work is an inventory of where your systems use encryption, and a plan to update it as your vendors adopt the new standards. That is maintenance, not a prediction.

How should a small business prepare for AI agents in 2027?
The original post ended with a long checklist for business leaders, professionals and policymakers. Most of it was sound but vague. Here is the shorter version I would give a business owner today, built on the guidance above.
Give AI agents less access than they ask for

Start with an inventory: which AI agents can read your email, files, calendar, CRM or bank portal, and with whose login. Then cut. OWASP names "excessive functionality," "excessive permissions" and "excessive autonomy" as the root causes of what it calls excessive agency. (OWASP LLM06:2025 Excessive Agency (opens in a new tab)) A read only connection is often enough. A separate account for the agent, with only the folders it needs, limits what a hijacked agent can reach. This is the same idea as zero trust, applied to software that acts for you.
Keep a person on the approve button

OWASP's advice is direct: "Utilise human-in-the-loop control to require a human to approve high-impact actions before they are taken." (OWASP LLM06:2025 Excessive Agency (opens in a new tab)) Decide in advance what counts as high impact for your business. Sending mail to a customer, paying an invoice, deleting records and changing a password are good places to start.
Log what AI agents do

If you cannot see what an agent did, you cannot tell a mistake from an attack. The NCSC suggests logging enough to spot suspicious activity, "potentially including full input and output of the LLM, as well as tool use, API calls etc." (NCSC, December 2025 (opens in a new tab)) For a small team, that can be as simple as a weekly look at what each agent sent, changed or bought.
Write the rules down
A one page policy beats a perfect one nobody reads: which AI agents are approved, what data they may touch, who approves their actions, and who to call when something looks wrong. Our knowledge base has a plain guide to AI agent governance for a small business, and I wrote about governing the sequence of agent actions, not each click (opens in a new tab).

Frequently asked questions
Did AI agents replace web browsers in 2026?
No. AI agents now act inside browsers, email and business apps, and use of AI chatbots reached about half of U.S. adults in Pew's 2026 survey. Most people still browse and search the usual way.
Is prompt injection fixed yet?
No. OWASP says it is unclear whether fool-proof prevention exists, and the UK NCSC says it may never be fully mitigated. The practical answer is to limit what an agent can do if it is fooled.
Does Massachusetts have an AI law for businesses?
Not a dedicated one as of October 2026. The Attorney General's 2024 advisory says existing consumer protection, anti-discrimination and data security laws apply to AI. A state data privacy bill is in conference committee.
What is the first step to using AI agents safely?
List every AI agent that can reach your accounts and data, and the login it uses. Then remove access each one does not need for its task.
Should an AI agent ever act without a person approving?
For low risk, reversible tasks, often yes. For anything that sends, pays, deletes or changes access, OWASP recommends a person approves the action first.
This article is general information, not legal advice. Kief Studio is not a law firm, and reading it does not create an attorney-client relationship. Laws and regulations change, so check the current text and talk with a licensed Massachusetts attorney about your situation.
Where we fit

Kief Studio is in Shrewsbury, and we help businesses across Massachusetts put AI agents to work without handing them the keys. Our AI implementation work takes an idea out of the demo and into daily work, with quality gates on every output and a human path for anything the model is unsure about. When a job has several steps, our multi-agent workflows give each agent one task, only the tools that step needs, and a log of every action. We run it for you on LTFI, so nothing is installed on your side.
On the security side, managed security monitoring means we take inventory of what runs and who can reach it, keep it patched, and report what changed in plain language. We do not build agents to take the place of your people. Your people drive the work, and they approve the decisions that need a person.
First conversation is free. No commitment.




