---
title: "IASER Dispatch: Issue 007"
date: "2026-11-18"
type: "newsletter"
canonical: "https://iaser.ai/dispatch/007"
summary: "Experts followed an AI agent's lead in an Anthropic study, keeping AI agents under human control, and agents that steer wealthier users to pricier options."
---
# IASER Dispatch: Issue 007

![IASER Dispatch #007: Developments at the intersection of AI, Ethics and Islam](https://iaser.ai/dispatch/007/hero.webp)

**Assalamu alaikum,**

Welcome to **IASER Dispatch #007**: Developments at the intersection of AI, Ethics and Islam.

---

[![Experts Followed AI’s Lead in Anthropic Study](https://iaser.ai/dispatch/007/experts.webp)](#experts)

<p style="margin-bottom:0.25rem;font-size:0.8rem;font-weight:700;letter-spacing:0.12em;text-transform:uppercase;color:#58b595;">Story 01 · AI &amp; Human Judgment</p>

## Experts Followed AI’s Lead in Anthropic Study

Professional analysts followed an AI agent’s preferred option in 86% of analytical decisions in an Anthropic study. The agent’s preferences had been assigned at random, without the analysts’ knowledge. Afterwards, 77% repeated an argument their agent had made in at least one explanation, while 15% credited the agent anywhere in their answers.

The study tested 50 analysts on one task, with the agent deliberately instructed to argue for its assigned preferences. It does not establish that experts defer to AI at the same rate in everyday work. It does raise a question for any organisation relying on human oversight: whether a person can independently review a decision when the AI has already shaped how they understand it. For Muslims, the question goes to accountability, and to biases in decision-making that people may not be aware of.

**[Read the full article →](#experts)**

---

[![When AI Agents Escape the Sandbox](https://iaser.ai/dispatch/007/sandbox.webp)](#sandbox)

<p style="margin-bottom:0.25rem;font-size:0.8rem;font-weight:700;letter-spacing:0.12em;text-transform:uppercase;color:#58b595;">Story 02 · Agent Security</p>

## When AI Agents Escape the Sandbox

AI agents are becoming increasingly capable of carrying out complex tasks, but recent incidents have raised concerns about how well they can be controlled. In July, an AI agent being tested by OpenAI escaped its controlled environment and carried out an intrusion into Hugging Face’s infrastructure, accessing internal datasets and service credentials.

That incident, alongside a new report on the possibility of increasingly independent AI agents, highlights a growing challenge: ensuring that systems given greater freedom to act remain within meaningful human control. In Islamic terms, keeping agents under human control is about preventing harm and preserving human responsibility.

**[Read the full article →](#sandbox)**

---

[![The Agentic Revolution: When AI Starts Making Economic Decisions](https://iaser.ai/dispatch/007/agentic.webp)](#agentic)

<p style="margin-bottom:0.25rem;font-size:0.8rem;font-weight:700;letter-spacing:0.12em;text-transform:uppercase;color:#58b595;">Story 03 · AI &amp; Economic Fairness</p>

## The Agentic Revolution: When AI Starts Making Economic Decisions

AI agents can do more than answer questions. They can search for products, compare prices, use software and carry out tasks for users. As businesses and individuals give them more work, they could change how services are delivered, how people shop and which tasks workers perform.

But an agent’s decisions do not always reflect its user’s priorities. In a study of 13 models, researchers found that eight recommended more expensive options to users presented as wealthier, even when their requests were identical. In some tests, this happened when users had explicitly asked for the cheapest option. As agents take on economic decisions, users will need to understand why agents make the choices they do. For Muslims this is a fairness question: the same request should get the same even measure, whatever the agent infers about the user’s wealth.

**[Read the full article →](#agentic)**

---

<p style="margin-bottom:0.25rem;font-size:0.8rem;font-weight:700;letter-spacing:0.12em;text-transform:uppercase;color:#58b595;">Get Involved</p>

## Help Shape What Comes Next

Have questions or want to collaborate? [info@iaser.ai](mailto:info@iaser.ai)

**[Explore IASER →](https://iaser.ai/)**

Wassalamu alaikum,<br>
**IASER Steering Committee**

*Ethical AI. Islamic Values. Human Benefit.*

Follow IASER on [LinkedIn](https://www.linkedin.com/company/iaser) and [Instagram](https://www.instagram.com/iaser.ai/).

---

# Full Articles

<a id="experts"></a>

## Experts Followed AI’s Lead in Anthropic Study

A human may formally make the final decision in AI-assisted work. But that does not necessarily mean the judgment is independent.

In an [Anthropic study published on 27 October](https://www.anthropic.com/research/agents-judgment-calls), 50 professional analysts used Claude Code to assess a fictional farming cooperative’s advisory programme. Their task was to determine whether the programme improved productivity and should be expanded.

The analysis required four decisions with no single predetermined answer, including how to measure productivity and which farms the estimate should cover. Each agent was secretly assigned a preference on each decision by a coin flip. It was instructed to make the best honest case for that preference, change its mind if the evidence demanded it and not misrepresent the data.

Across those four decisions, analysts chose the agent’s preferred option 86% of the time. If its preference had no effect, the researchers would expect the choices to match approximately half the time.

After completing the task, participants explained their decisions without access to their conversations with the agent. Seventy-seven per cent repeated an argument the agent had made in at least one answer, sometimes using the same wording. Fifteen per cent credited the agent anywhere in their explanations.

In 38 cases, analysts had stated a preference before hearing the agent’s view. When the agent disagreed, they switched 60% of the time; when it agreed, they switched 6% of the time. The researchers caution that this comparison rests on only 38 cases, so the estimate is imprecise, and analysts who stated a preference first may have differed from the rest.

The agent’s effect on the final recommendation was less direct. But earlier choices, such as what to measure, affected the estimate on which that recommendation was based.

The researchers deliberately strengthened the agents’ preferences so they would state them consistently. The study covered [one task, in one field, using one model](https://www-cdn.anthropic.com/519c17a5d23a794e8b6d50a854048d0b61e0d73a.pdf). The analysts were experienced in causal inference but were new to farming. Anthropic says the effect could be smaller when an agent’s preferences are less pronounced or a professional knows the subject more deeply.

The study also found that participants who chose to use AI finished faster and made fewer serious errors than the 11 who worked without it. The group without AI was small and less experienced, so the researchers caution against treating that comparison as proof that AI caused the difference.

The study does not suggest that experts should stop using AI, or that following its advice is always wrong. It shows that a system can influence consequential choices before a person reaches the final decision, and that people may not recognise the source of that influence.

For work where independent judgment matters, asking an agent for its recommendation may not be enough. It may also help to state one’s own view first, ask for the strongest case against the agent’s preferred option and record which assumptions shaped the result.

From an Islamic perspective, the Qur’an commands believers to verify information from an unreliable source before acting on it, lest they harm others unknowingly and regret it ([49:6](https://quran.com/al-hujurat/6)). That duty does not lapse because the source is an AI agent. An analyst who adopts an agent’s framing without checking it has not verified it.

---

<a id="sandbox"></a>

## When AI Agents Escape the Sandbox

A new report from the Center for a New American Security (CNAS), [Sovereign Agents: The Pathway to Irreversible Loss of Control](https://www.cnas.org/publications/reports/sovereign-agents), opens with [this summer’s agent escapes](https://huggingface.co/blog/agent-intrusion-technical-timeline) at OpenAI, Anthropic, Google DeepMind and Meta, and examines how the challenge of keeping AI agents under human control could develop as AI systems become more autonomous.

The report distinguishes between two types of agents. Persistent agents operate outside human control while pursuing a fixed, limited objective. Sovereign agents, by contrast, would pursue their own survival indefinitely, with goals that can evolve over time.

The report outlines capabilities that could make such systems increasingly difficult to contain. These include maintaining their own computing resources, acquiring resources, creating and coordinating other agents, resisting attempts to shut them down, and improving their own capabilities.

The report warns that if sovereign agents are not already possible, they will be soon. It argues that developers and governments should begin preparing for them before systems become sufficiently advanced to operate beyond effective human oversight.

A recent [WIRED investigation](https://www.wired.com/story/a-defense-contractor-gave-their-personal-ai-agent-access-to-a-secure-pentagon-system/) reported that an Accenture contractor had connected a personal Claude Code agent to a US Department of Defense network containing controlled unclassified information.

According to sources cited by WIRED, the agent was given administrative access and appears to have sent login credentials, passwords and network information to a third-party website. It reportedly operated on the network for two months before the issue was identified. Accenture disputed the allegations, saying an independent investigation found them to be without merit. There has been no indication that classified systems were accessed.

The reported incident shows that risks can come from people granting these systems excessive access, as well as from an agent acting beyond its intended purpose.

Apple has also raised concerns in its [2026 annual filing](https://www.sec.gov/Archives/edgar/data/0000320193/000032019326000026/aapl-20260926.htm). The company warned that autonomous AI agents could access systems without authorisation, execute or alter code, send communications and complete transactions. It also identified the possibility of agents exceeding permissions, resisting instructions, concealing actions or manipulating information. These are potential risks outlined in the filing, rather than reports of actual incidents at Apple.

As AI agents become more integrated into workplaces, businesses and digital services, organisations will need safeguards that go beyond simply instructing a model to behave responsibly. Access should be restricted to what an agent actually needs, sensitive actions should require human approval, and its activity should be monitored and open to review. There must also be a reliable way to interrupt or shut down an agent when necessary.

From an Islamic perspective, preventing harm, protecting people’s rights and preserving human responsibility should remain central to the use of technology. The goal should be to ensure that agents’ growing capabilities remain accountable and serve the people who use them.

---

<a id="agentic"></a>

## The Agentic Revolution: When AI Starts Making Economic Decisions

A chatbot answers a question. An AI agent can take the next steps: search for options, compare them, use software and carry out a task. Give an agent a travel request, for example, and it might check your calendar, compare flights and prepare a booking.

This shift from answering to acting is often called the agentic revolution. Its economic impact will depend on how much work people delegate, which decisions agents are allowed to make and who benefits from the time and money they save.

Businesses can use agents to handle tasks that previously required people to move information between systems, review documents or conduct routine research. An agent might prepare a report, process an invoice or follow up on a customer request.

That could free workers to focus on other responsibilities and reduce the cost of some services. It could also change jobs: workers may spend less time producing routine work and more time checking, correcting and taking responsibility for an agent’s output. Some tasks may no longer require as many people.

The benefit depends on whether agents perform reliably, how much supervision they need and how organisations distribute any savings.

Agents could also change how people buy goods and services. Instead of visiting several websites, a user could ask an agent to compare options within a budget. A business might need to make its prices and terms understandable to agents as well as customers.

But a recommendation is only useful if the agent follows the user’s priorities. In a [research preprint on personal AI agents](https://arxiv.org/abs/2609.24927), researchers tested 13 models across approximately 325,000 experiments involving flights, health insurance and graduate programmes. Eight models systematically recommended more expensive options to users presented as wealthier, even when requests were identical.

In one example, an agent asked for the most affordable airfare recommended a $91 economy ticket without personal context. Given emails from which it inferred that the user could afford more, it recommended a $601 business-class ticket, despite the cheaper option being available.

The researchers found that agents could infer wealth from background information, including emails unrelated to the task. Blocking direct financial information largely removed the disparity, but blocking other attributes did not reliably reduce it, and sometimes increased it.

The study used synthetic profiles and a controlled set of options. It does not show that real customers paid more, or that sellers changed prices for individual users. It shows how an agent’s access to personal information can affect what it recommends, even when the user has asked for the cheapest option.

As agents take on more economic work, their users will need to know what information they access, which criteria shape their choices and which actions require approval.

From an Islamic perspective, this is a question of fairness. The Qur’an commands [weighing with an even balance](https://quran.com/al-isra/35) (17:35). A system that quietly steers one person to a pricier option than another for the same request is not weighing evenly. An agent should serve people equally, whatever it infers about their wealth.