Skip to content
Apart Research
Apart Research
Contact us
Sprint projectMar 22, 2026Cape Town

When Safety Becomes the Vulnerability

Caleb Rudnick, August Lina · Team augusta_caleb

Submitted to AI Control Hackathon 2026. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: When Safety Becomes the Vulnerability

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

Any base64-encoded string included in a message to the Claude API causes that request to fail. This behaviour is reproducible across Sonnet and Opus 4.6, and across all platforms including the API, web and mobile applications, and Claude Code. While the blanket rejection of base64 content was originally a reasonable defence against prompt injection—attackers could encode malicious instructions to bypass keyword-based safety filters—the measure has become a liability as large language models have moved from conversational assistants to critical infrastructure components.

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. the report is very well written and presented to the level of a good conference paper.

    im a bit skeptical of the threat model (base64-injection attacks causing critical system failures due to hitting refusals from the model provider). I feel like systems should be basically robust to model outages and base64 attacks might be a specific cause of outages. My guess is there might be narrow situations in which an attack can strategically cause a model outage leading to a security vulnerability, but that this isn't a super important threat vector.

    FYI I tried running a simple example from the report (aGVsbG8gd29ybGQ= “hello world”) in claude.ai and it did not trigger a refusal (the model just understood the message and responded appropriately).

    I agree with the broad dynamic:

    >This is a concrete instance of a broader principle in AI control research: that safety mechanisms which

    fail catastrophically (by crashing the system or returning no response) rather than degrading

    gracefully (by flagging suspicious content and continuing) can be weaponised by adversaries

    who understand the failure mode.

    But I think this type of threat should just be caught by standard testing and security protocols and isn't a major threat model. I think it's good the report points out the dynamic.

    Read full reviewShow less

Cite this project

@misc{rudnick2026safety,
  title = {{When Safety Becomes the Vulnerability}},
  author = {Caleb Rudnick and August Lina},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/when-safety-becomes-the-vulnerability-gi2u}},
  url = {https://apartresearch.com/sprints/projects/when-safety-becomes-the-vulnerability-gi2u}
}

More from AI Control Hackathon 2026

All 125 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026