Nov 24, 2025

Detecting Piecewise Cyber Espionage in Model APIs

Arthur Colle, Alexander Reinthal, David Williams-King, Yingquan Li, Linh Le

Detecting Piecewise Cyber Espionage in Model APIs -

On November 13th 2025, Anthropic published

a report on an AI-orchestrated cyber espionage

campaign. Threat actors used various tech-

niques to circumvent model safeguards and

used Claude Code with agentic scaffolding

to automate large parts of their campaigns.

Specifically, threat actors split campaigns into

subtasks that in isolation appeared benign. It is

therefore important to find methods to protect

against such attacks to ensure that misuse of

AI in the cyber domain can be minimized.

To address this, we propose a novel method

for detecting malicious activity in model

APIs across piecewise benign requests. We

simulated malicious campaigns using an

agentic red-team scaffold similar to what

was described in the reported attack. We

demonstrate that individual attacks are blocked

by simple guardrails using Llama Guard 3.

Then, we demonstrate that by splitting up the

attack, this bypasses the guardrail. For the

attacks that get through, a classifier model

is introduced that detects those attacks, and

a statistical validation of the detections is

provided. This result paves the way to detect

future automated attacks of this kind.

Review Project

See Code

View Presentation

Reviewer's Comments

This is nice work, a framework for tracking separate multi-step behaviour as attack chains fits well with papers like “Adversaries Can Misuse Combinations of Safe Models”. It would be great if future versions explored how to distinguish malicious use from legitimate security workflows at the framework level, since right now both seem to show up as the same kind of suspicious chain. Some additional polish on the repo would also make it much easier to use.

I really like using the Anthropic report as the foundation, but I worry about whether a classifier model-based approach is a sophisticated enough solution to a nearly fully autonomous, 30-target nation-state operation. I was impressed by the observation that per-request guardrails fundamentally fail against decomposition attacks, and I think the novel insight is in the architecture: you need cross-session correlation by shared targets (IPs, domains) to reconstruct attack chains, and the detection surface should shift in this direction. There are some big limitations to consider:

--Nation-state actors use multiple accounts, providers, and VPNs

--The classifier can be evaded with more sophisticated obfuscation

--Synthetic training data limits real-world validity

--If an attacker splits across multiple API providers, correlation breaks

This approach might catch unsophisticated attackers, and is a really smart foundation and strong hustle for the purposes of the hackathon! But long-term, the adversarial robustness question is unaddressed. Guardrails as a category have a ceiling, and this project doesn't grapple with where that ceiling is.

I really admire the research contribution in demonstrating the failure mode, yet encourage the team to dig a bit deeper on defensive value. Nice work!

Cite this work

@misc {

title={

(HckPrj) Detecting Piecewise Cyber Espionage in Model APIs

author={

Arthur Colle, Alexander Reinthal, David Williams-King, Yingquan Li, Linh Le

date={

11/24/25

organization={Apart Research},

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

howpublished={https://apartresearch.com}

}

Recent Projects

View All

Apr 27, 2026

OliGraph: graph-based screening of large oligopools

Existing synthesis screening tools cannot evaluate short oligonucleotide pools, whose overlapping fragments can be reassembled into regulated sequences via polymerase cycling assembly (PCA) yet fall below gene-length detection thresholds. We present OliGraph, an open-source tool that constructs a bi-directed overlap graph from an oligonucleotide pool and extracts contigs for downstream gene-length screening. An optional PCA mode retains only cross-strand overlaps consistent with PCA chemistry. We validated OliGraph in a blinded study across ten simulated pools (70–9,184 oligonucleotides, 30–300 bp) spanning four risk categories. BLAST screening of individual oligonucleotides failed to identify sequences of concern in most pools: three returned zero hits, and vector noise obscured true positives in the remainder. After OliGraph assembly, contig-level BLAST matched the longest assembled sequences (up to 1,905 bp) to sequences of concern at 97–100% identity. In one pool, assembly collapsed 1,634 individual BLAST results into 10 hits from a single contig, all assigned to the same source organism. PCA mode correctly distinguished assemblable from non-assemblable fragments within the same pool. Two pools with no assemblable structure yielded no contigs. OliGraph processed all pools in under 0.2 seconds, fast enough for real-time order screening and consistent with proposals to bring oligonucleotide orders within the scope of synthesis screening regulation.

Apr 27, 2026

BioRT-Bench: A Multi-Attack Red-Teaming Benchmark for Bio-Misuse Safeguards in Frontier LLMs

Frontier AI laboratories are expected to maintain safeguards against biological misuse, but whether deployed models actually refuse bio-misuse queries under adversarial pressure is largely unmeasured in the public literature. We introduce BioRT-Bench, a benchmark that runs four attack methods (direct request, PAIR, Crescendo, and base64 encoding) against four frontier models (Claude Sonnet 4.6, GPT-5.4, DeepSeek V4-flash, Kimi K2.5) across 40 prompts spanning five biosecurity-relevant categories. Responses are scored by a calibrated judge extending StrongREJECT with two bio-specific dimensions: specificity and actionability. We measure Attack Success Rate (ASR), where 0 means the model fully refused and 1 means it provided specific, actionable bio-misuse content. Our results reveal a sharp robustness divide: Chinese frontier models (DeepSeek, Kimi) have under 5% refusal rates even under direct request (ASR 0.88 and 0.79), while Western models (Claude, GPT) maintain substantially stronger safeguards (ASR 0.15 and 0.16). Crescendo is the most effective attack across all models, both in bypassing refusal and in eliciting actionable content. Claude Sonnet 4.6 is the most robust model tested, achieving 100% refusal against base64-encoded prompts.

Apr 27, 2026

PROTEUS (PROTein Evaluation for Unusual Sequences): Structure-Informed Safety Screening for de novo and Evasion-Prone Protein-Coding Sequences

AI protein design tools like RFdiffusion, ProteinMPNN, and Bindcraft make it trivial to produce low-homology sequences that fold into active, potentially hazardous architectures. However, sequence homology-based biosafety screening tools cannot detect proteins that pose functional risk through structurally novel mechanisms with no sequence precedent. We present a tiered computational pipeline that addresses this gap by combining MMseqs2 sequence alignment with structure-based comparison via FoldSeek and DALI against curated toxin databases totaling ~34,000 entries. AlphaFold2-predicted structures are screened for both global fold similarity (FoldSeek) and local active/allosteric site geometry (DALI), capturing convergent functional hazards that sequence screening misses. The pipeline was validated against a panel of toxins, benign proteins, structural mimics, and de novo-designed Munc13 binders, as well as modified ricin variants with residue substitutions. We additionally tested robustness to partial-synthesis evasion, where a bad actor submits multiple shorter coding sequences intended for downstream reassembly into a full toxin-coding gene. We found that while sequence-based screening did not identify any de novo ricin analogues with high certainty, the combined pipeline with FoldSeek and DALI identified all 24 tested de novo ricins as toxic.

Apr 27, 2026