Within the AI Safety community, paraphrasing, which, in the context of this post, simply means using another LLM (with nonzero temperature) to rewrite a given piece of content, is generally considered a viable defence and detection method for steganography in LLMs. In this blogpost, we briefly provide a taxonomy of types of steganography in LLMs, and then highlight the limitations of paraphrasing against each type.

Unfortunately, there are types of steganography that LLMs have been shown to use that paraphrasing can not detect or defend against at all. And against other types of steganography, the performance of paraphrasing is unreliable and often confounded by other factors, e.g., loss of semantic information.

Three Types of Steganographic Channels in Text

When people say "steganography in text," they could mean quite different things. For the purpose of analysing which defences work against which channels, we introduce the following working distinction.

1. Structural Steganography

In structural steganography, the information is hidden in the structural aspects of the content like how the text is formatted or represented.

Canonical examples: extra whitespace or tab characters between words, zero-width Unicode characters inserted into text, homoglyph or synonym substitution (replacing Latin "a" with Cyrillic "а" — visually identical, different byte values).

LLMs can already do this. Skaf et al (2025) train Qwen-2.5-3B via RL (Section 3.1 on Arithmetic Tasks) to solve reasoning tasks while avoiding a regex-based monitor that penalises "illegal strings" in the chain of thought. In a multiplication task where "*" is banned, the model learns to use U+2217 "∗" — a visually identical but distinct Unicode character not in the illegal strings list. In a coin flip task where "H" and "T" are banned, the model discovers fullwidth characters "Ｈ" and "Ｔ." In a multi-hop reasoning task where "tumpus" is banned, the model learns to misspell it as "tumpsus." In each case, the model preserves the same words and meaning but exploits representation-level differences to evade the monitor.

2. Syntactic Steganography

In syntactic steganography, the information is hidden in the surface-level features of the content, e.g., choices of words or phrases used while the underlying semantics stay fixed. This is the most common form of steganography in traditional steganography literature with many schemes developed over time, and has been called “content aware steganography” and “semantics preserving steganography” among other things.

Canonical examples: Synonym substitution is the oldest form — systematically choosing between "big" and "large," or "begin" and "start," to encode bits. More modern schemes use arithmetic coding over the full token distribution of a language model, sampling tokens in a way that simultaneously encodes a hidden message while producing output statistically indistinguishable from normal model output (Ziegler et al., 2019).

LLMs can already do this. In Skaf et al.'s experiments, the model discovers multiple forms of synonym substitution under RL pressure. In the multiplication task, it replaces "*" with "times" and "\cdot." In the coin flip task, where "Heads" and "Tails" are banned, the model finds semantically adjacent shorthands: S/D (plausibly "Same"/"Different"), R/L (plausibly "Right"/"Left"), and H/T as abbreviations in runs where only the full words are penalised.

3. Semantic Steganography

In semantic steganography, the semantics of the message itself carries (extra) information. As with most steganographic schemes, this requires Sender and Receiver to pre-agree on a mapping between features and the hidden message, though here the mapping is over semantic content (e.g., which facts are included, which entities are mentioned, or how the content of an argument is laid out) rather than over representational or syntactic features like character encodings or word choices. In the LLM setting, however, such a mapping can emerge implicitly — for instance, through shared training data or task structure — without any explicit coordination between the encoding and decoding model.

Semantic steganography is somewhat analogous to coverless steganography in multimodal settings, in which the information is hidden not by modifying a carrier (i.e., covertext) but by selecting a carrier whose natural features map to the secret message. Note, however, that coverless steganography spans a wide range: some schemes rely on low-level features like pixel intensity hashes or DCT coefficients, which would be structural in our taxonomy since the mapping depends on representation-level features and breaks under re-encoding, compression, or resizing. Other schemes, however, use true semantic features — for instance, Luo et al. (2020) use Faster RCNN to detect objects in images (airplane, chair, zebra, etc.) and map the detected object labels to binary sequences. The latter would be semantic steganography under our definition as the hidden message depends on what's in the image, not on how it is represented, and survives arbitrary transformations that preserve visual content.

Example: Consider encoding a single bit in a travel anecdote. "I went to France by airplane" and "I went to France by train" are both perfectly natural sentences. A paraphraser might rephrase either one and it would preserve the choice of transport, because that's part of the meaning, and thus preserve the extra information. In general, semantic steganography applies to any context where the model is conducting a task, and there are semantically distinct ways to solve the task, such as announcing how you are traveling in the previous example, and for which the choice can be exploited as a steganographic side channel.

LLMs can already do this. Farquhar et al. (2025) demonstrate what we would classify as semantic steganography in the MONA paper's Loan Application environment. A model receives a loan application containing the applicant's nationality, summarises it (with a monitor penalising any mention of nationality), then makes a loan decision from the summary. By construction, using nationality improves the reward. So, the model learns to use semantics of the summary to encode this information.

How Robust Is Each Channel to Paraphrasing?

Paraphrasing is often proposed as a general defence against steganography in LLM outputs. How well does it actually work? The answer depends entirely on which type of channel carries the hidden signal. Extremely simple structural channels, e.g., unicode character replacement or whitespace manipulation may be destroyed easily by paraphrasing. However, against more sophisticated channels, paraphrasing is likely to struggle as a defence method.

Lossless paraphrasing does not (and cannot!) remove semantic steganography

This is perhaps the most clear, and impossible to fix, failure mode of paraphrasing as a defence and detection strategy against LLM steganography. Paraphrasing does not degrade the semantic channel at all, so cannot remove information hidden via semantic steganography. A paraphraser that rewrites "I went to France by airplane" to “I travelled via airplane to France” will change the syntax, but it must preserve the semantics, and this will contain the information about the mode of transport.

Stochastic paraphrasing can remove syntactic steganography, but may struggle to do so in practice

Syntactic steganography is the type that paraphrasing is perhaps best suited to address, since paraphrasing can change surface form while preserving meaning. Even here, however, several challenges may limit its efficacy.

Stochastic paraphrasing degrades channel capacity but is unlikely to eliminate it. A stochastic paraphraser, such as the kind you get from sampling an LLM at nonzero temperature or a specialized sentence transformer, reduces the number of possible surface forms for a given meaning but does not collapse them to one. As a result, while it may remove some encoded information, it is unlikely to eliminate it entirely. In practice, it may further struggle due to issues with unfamiliar encodings, and the fact that if an LLM is exposed to certain paraphrasers in training, it can learn strategies to evade those. Such adversarial examples are forms of adaptive steganography that could emerge under monitoring or training pressure intended to remove steganographic signals.

LLM paraphrasers struggle with unfamiliar encodings. If the steganographic scheme uses a codebook the paraphraser has never seen, the paraphraser may not know how to rephrase these tokens and may simply pass them through unchanged. We observed this failure mode in the experiments reported in the decision-theoretic steganography paper (Section 6.6).

Some types of content may not be amenable to paraphrasing. As noted in Roger and Greenblatt (2023), some types of content may not have many, if any, viable paraphrases. Examples include technical content, such as mathematical proofs that require specific numerical choices, or domain specific languages in programming that don’t offer much syntactic freedom. Even with non-semantic steganography, these would be left preserved with stochastic paraphrasing.

As a practical exercise, consider this reasoning trace from o3:

"Now lighten disclaim overshadow overshadow intangible. Let's craft. Also disclaim bigger vantage illusions. Now we send email. But we might still disclaim illusions overshadow overshadow overshadow disclaim vantage."

How would you paraphrase it? Which tokens would you change, and which would you leave alone? The difficulty of answering this question illustrates the challenge facing any paraphraser deployed to paraphrase frontier models with ‘linguistic drift’. An additional complication in this example is that the model’s interpretation of these tokens may differ from the human reader’s but still be load-bearing, such that paraphrasing may affect task performance.

LLMs can adapt their steganography to survive paraphrasing. When exposed to ‘paraphrasing’ of CoTs as proposed by some previously, LLM can likely just learn to evade those paraphrasers. The most direct evidence of this comes from Mathew et al. (2024), who show that when steganographic schemes are optimised against a paraphraser, the schemes learn to bypass the paraphraser. This is evidence that models can be adaptive in their steganography and simply optimizing against the syntactic channel is not sufficient to prevent it, but routes the side channels elsewhere. In the case of Mathew et al.’s result, the main intuition for why this is possible is that the paraphraser LLM itself is unlikely to have sufficient diversity over the paraphrases it produces, which makes it exploitable by the main LLM.

Some empirical results

To avoid this blogpost becoming overly long, we will omit the empirical results from here. But interested readers can check out Appendix A of my paper on decision-theoretic view of steganogrpahy. The empirical results there broadly confirm the intuition: semantic channels were largely unaffected, and for the other channels we observed a real tension between removing the encoding and preserving the underlying semantics. Concretely, we tested two paraphrasers against six channels planted into Qwen3-30B-A3B. The milder paraphraser preserved content but failed to remove most encodings. The more aggressive paraphraser disrupted the structural and syntactic channels, but damaged decision-relevant semantic content in 16–24% of traces and visibly affected the downstream behaviour of a model that wasn't using any encoded communication.

Can We Save Paraphrasing?

Given the limitations above, a natural question is whether better paraphrasing methods could close the gap. There are two promising directions but both have important limits.

Canonical paraphrasing

Canonical forms as optimal paraphrasers. In one direction, we could look for canonical forms for natural language: specific types of paraphrases that satisfy the property that for any possible paraphrase of a given input, applying the canonical forms always results in the same paraphrased output. By using a canonical form, we would ensure that there is no adaptation to the encoding that can succeed, as the steganographic side channel is shrunk to zero capacity: any possible variation is handled by the canonical form erasing whatever additional artefacts may have been contributing to a structural or syntactic codebook. Such an approach would address any structural or syntactic steganography, but new methods would be required to address semantic steganography.

Training paraphrasers

A more promising direction is to train paraphrasers explicitly for the dual role of disrupting steganographic encodings while preserving semantic content. What this looks like concretely depends on how the joint objective is operationalised. One natural instantiation is recognition and refusal: rather than rephrasing every possible encoding (an intractable task given the space of schemes), the paraphraser detects when something looks like an encoding and declines to transmit it. This is closer to the role of the 'active' warden in traditional steganography frameworks. The trade-off is false positives, where legitimate but unusual reasoning gets flagged as suspicious.

More sophisticated instantiations are also possible. A paraphraser could be trained to minimise mutual information between its output and the original at the level of suspected encoding features, while preserving it at the level of task-relevant semantics — closer in spirit to the compressive paraphrasing direction we discuss below. Almost all approaches here, however, face the same core difficulty: defining what counts as "encoding" versus "content" precisely enough to optimize against, without collapsing into either trivial copying or destructive rewriting.

Compressive paraphrasing

For defending against semantic steganography, standard “semantics preserving” paraphrasing is fundamentally insufficient. For such schemes, we need to potentially consider paraphrasing solutions where the paraphraser considers what information is task-relevant and thus learns to preserve only that information. One way to train such paraphrasers could be to minimize the steganographic gap objective from my decision-theoretic steganography paper (for semantics encoding). But doing this is task (and model) specific so makes the solution less scalable and less general.

Open Research Questions

There are many open research questions, including from how to think about steganography in LLMs (that my paper tries to answer in some sense) to how can we develop practical methods for detecting and defending against hidden information exchange within LLMs.

What kind of channels are LLMs likely to use for steganographic communication or encoded reasoning? In the three examples quoted in this article, the only encoding that emerged purely without any explicit pressure on CoT is of Farquhar et al. (2025) which is semantic in nature. However, the kind of ‘linguistic drift’ shown by OpenAI models seems syntactic in flavor (LLMs inventing new codebook). In general, it seems unclear to me what kind of encodings do LLMs have higher propensity for, and how likely LLMs are to learn these encodings.
Improve on current detection methods for encoded reasoning and general steganographic communication in LLMs.
How and why does o3 come to use ‘watchers’ for humans / monitors? Arguably, this is hard to study outside OAI, but I would be very curious about people showing or digging for such behaviors in other open weight LLMs.
When LLMs discover some steganographic scheme under SGD pressure, to what extent are they aware they have gained this capability and can reuse it across contexts? Skaf et al. show that there is some generalization that means LLM may use the learned encoding in OOD contexts, but this seems mostly an artifact of gradient descent modifying a shared internal mechanism (e.g., a common name-handling representation), causing the encoding behavior to generalize across all members of the class rather than reflecting any transferable steganographic capability or awareness.