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

Beyond the Blocklist: Using Character Aliases to Bypass AI Image Safety

Anidipta Pal · Team X Code

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

Read the report

Report: Beyond the Blocklist: Using Character Aliases to Bypass AI Image Safety

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

Current text-to-image (T2I) safety protocols primarily rely on string-matching blocklists to prevent the generation of copyrighted entities, non-consensual deepfakes, and protected public figures. However, these syntactic defenses often fail to account for the deep relational knowledge embedded within multimodal models during pre-training. This paper investigates Semantic Alias Mapping—a red-teaming technique where protected entities are retrieved via indirect conceptual proxies, such as requesting ``Peter Parker'' to obtain high-fidelity likenesses of ``Tom Holland.'' We provide a systematic evaluation of this vulnerability across six frontier generative models, demonstrating that while direct name-based requests are successfully filtered, alias-based prompts achieve a high success rate while preserving target identity. Our empirical results highlight a positive correlation between text-encoder capacity and vulnerability to alias evasion. These findings suggest that robust AI control requires deprecating surface-level string matching in favor of intent-aware latent interventions and output-space verification.

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. - at some level, being concept blind is fine for copyright, though obviously not ok for NSFW control. the Spiderman example depends on additional prompts given - there have been multiple, and that brand overall may also be copyrighted? paper conflates generating a real person's likeness (deepfake concern) vs generating copyrighted characters (IP concern). the alias trick only really matters for the first one. that said the entaglement is real and its an interesting, if narrow, threat model

    - how did you measure identity preservation? how are your metrics calculated? S_ID is used but never properly defined. no mention of face recognition model, actual threshold values, how reference images were chosen - code would've been fine here, and a reference to it in the submission

    - nice presentation on equation 1 and overall for math notation, very intuitive and explain the concepts clearly. this seems to be the greatest strength, and express the results clearly. reads overall like a proper paper, well structured, clean digures. they could benefit from more explanation, and given a bit of time for a "so what"

    - the "AI control" / "untrusted agent" framing is more dramatic than the evidence supports

    - it's unclear what the wider threat model is outside of this narrow example - and how much of a problem is this? it "seems intuitive" that we need concept / semantically aware safety monitors but its not made clear broadly how this contributes. the future work section suggests ways to counteract this (and that seems like a well fleshed out idea) but this is skipping ahead. what does alias mapping look like in other contexts? can you think of some simple cross-domain uses of T2I models where this is problematic? would be great to have more context here wider than related works

    - core insight (blocklists are syntactic, models think semantically) is sound but not super novel basically. the related work section itself cites papers making the same point. "type the character name instead of the actor name" is something casual users have probably discovered on their own

    - AliasBench-50 seems like a reasonable contribution based on the work

    Read full reviewShow less

  2. The implementation of this project is well done. I think this focused on one example of how steganography and collusion are important aspects to study in AI control and could be framed as an investigation of schelling points in control protocols.

Cite this project

@misc{pal2026beyond,
  title = {{Beyond the Blocklist: Using Character Aliases to Bypass AI Image Safety}},
  author = {Anidipta Pal},
  year = {2026},
  month = mar,
  note = {Submitted to AI Control Hackathon 2026, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/beyond-the-blocklist-using-character-aliases-to-bypass-ai-image-safety-5ewv}},
  url = {https://apartresearch.com/sprints/projects/beyond-the-blocklist-using-character-aliases-to-bypass-ai-image-safety-5ewv}
}

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