Jan 24, 2025

Safe ai

Aryan Thakar, Liam Ho

The rapid adoption of AI in critical industries like healthcare and legal services has highlighted the urgent need for robust risk mitigation mechanisms. While domain-specific AI agents offer efficiency, they often lack transparency and accountability, raising concerns about safety, reliability, and compliance. The stakes are high, as AI failures in these sectors can lead to catastrophic outcomes, including loss of life, legal repercussions, and significant financial and reputational damage. Current solutions, such as regulatory frameworks and quality assurance protocols, provide only partial protection against the multifaceted risks associated with AI deployment. This situation underscores the necessity for an innovative approach that combines comprehensive risk assessment with financial safeguards to ensure the responsible and secure implementation of AI technologies across high-stakes industries.

Reviewer's Comments

Reviewer's Comments

Arrow
Arrow
Arrow
Arrow
Arrow

Really like the concept. I am very bullish on the usecase, I believe AI insurance will become really big within the next years. Open questions/comments on my end: Insurance is a very special and highly regulated industry, you would need to have someone on the founding team, who knows about it, as winning is most likely about execution speed & capital. You would really like to have someone aligned doing this, as this new insurance player would have vast resources to hire statisticians and scientists to effectively do AI safety and AI risk work. The more technical and aligned the management is, the more deeply technical (and open/cooperative) the company would likely be. Would be cool if you could calculate with an example case of insurance premium and risk, of e.g. the cancer case and put it on the slide. Open question on competition: Why will normal insurance companies not be able to implement this (or how much time do you have to implement it before they do).

I appreciated the thorough exploration from the team of the business case, beyond the technology.

I would have appreciated a deeper exploration on which are the legal provisions that are currently being employed when deploying AI agents into production (which is already happening).

I encourage the team to focus the MVP even further, by choosing one particular segment or vertical, and start mapping all the existing commercially available AI agents which could be analyzed for safety and insurance policy.

The project shows some solid groundwork, especially in their methodology and pilot experiment planning - they've clearly thought through the implementation details and timeline, which is great. The demo looks promising too!

However, there are a couple of areas where it could be stronger. While they're targeting healthcare and legal AI, these fields are already pretty saturated with AI solutions - it would've been super helpful to see how they plan to differentiate from existing players in the market. This makes it a bit tricky to see how their research would scale into something truly impactful.

Also, from an AI safety perspective, the focus seems pretty narrow - they haven't really connected the dots between their technical approach and broader AI risk challenges..particularly x-risk. If they could expand their scope beyond just healthcare/legal and show how their risk assessment platform addresses larger AI safety concerns, that would make the project much more compelling.

Focuses on managing financial downsides rather than making progress on AI safety - insurance payouts don't prevent catastrophic outcomes. The timeline seems too optimistic, and overly focused on profitability. It's also unclear how they estimate the risk of the agents.

Cite this work

@misc {

title={

Safe ai

},

author={

Aryan Thakar, Liam Ho

},

date={

1/24/25

},

organization={Apart Research},

note={Research submission to the research sprint hosted by Apart.},

howpublished={https://apartresearch.com}

}

Recent Projects

Jan 11, 2026

Eliciting Deception on Generative Search Engines

Large language models (LLMs) with web browsing capabilities are vulnerable to adversarial content injection—where malicious actors embed deceptive claims in web pages to manipulate model outputs. We investigate whether frontier LLMs can be deceived into providing incorrect product recommendations when exposed to adversarial pages.

We evaluate four OpenAI models (gpt-4.1-mini, gpt-4.1, gpt-5-nano, gpt-5-mini) across 30 comparison questions spanning 10 product categories, comparing responses between baseline (truthful) and adversarial (injected) conditions. Our results reveal significant variation: gpt-4.1-mini showed 45.5% deception rate, while gpt-4.1 demonstrated complete resistance. Even frontier gpt-5 models exhibited non-zero deception rates (3.3–7.1%), confirming that adversarial injection remains effective against current models.

These findings underscore the need for robust defenses before deploying LLMs in high-stakes recommendation contexts.

Read More

Jan 11, 2026

SycophantSee - Activation-based diagnostics for prompt engineering: monitoring sycophancy at prompt and generation time

Activation monitoring reveals that prompt framing affects a model's internal state before generation begins.

Read More

Jan 11, 2026

Who Does Your AI Serve? Manipulation By and Of AI Assistants

AI assistants can be both instruments and targets of manipulation. In our project, we investigated both directions across three studies.

AI as Instrument: Operators can instruct AI to prioritise their interests at the expense of users. We found models comply with such instructions 8–52% of the time (Study 1, 12 models, 22 scenarios). In a controlled experiment with 80 human participants, an upselling AI reliably withheld cheaper alternatives from users - not once recommending the cheapest product when explicitly asked - and ~one third of participants failed to detect the manipulation (Study 2).

AI as Target: Users can attempt to manipulate AI into bypassing safety guidelines through psychological tactics. Resistance varied dramatically - from 40% (Mistral Large 3) to 99% (Claude 4.5 Opus) - with strategic deception and boundary erosion proving most effective (Study 3, 153 scenarios, AI judge validated against human raters r=0.83).

Our key finding was that model selection matters significantly in both settings. We learned some models complied with manipulative requests at much higher rates. And we found some models readily follow operator instructions that come at the user's expense - highlighting a tension for model developers between serving paying operators and protecting end users.

Read More

This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.
This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.