Search: “longest-prefix-match”

31 stories

DomainAll Cyber AIWindow24h 3d 7d 30d AllIn the wild Important only Show duplicates

CategoryAll Exploit / PoC 1 Vulnerability 1

SeverityAll critical 0 high 1 medium 0 low 1 info 0

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

CVE-2026-86287: Net::IP::LPM versions before 1.12 for Perl accept malformed prefix lengths

CVE-2026-86287: Net::IP::LPM Perl library before 1.12 accepts malformed IP prefix lengths, risking incorrect longest-prefix-match results.

Net::IP::LPM versions before 1.12 for Perl accept malformed prefix lengths, per an oss-security disclosure posted September 7, 2026. The lax input validation in this longest-prefix-match library could produce incorrect matching behavior. The fix is available in version 1.12 on CPAN.

oss-securitySep 7, 18:40 UTC · 9d agoVulnerability CVE-2026-86287

BGP hijack infecting networks caused by a comedy of errors that’s not funny at all

Attackers BGP-hijacked Softaculous IP space via Hetzner routing lapses to push malicious Virtualizor updates to hosting servers.

Unknown attackers hijacked the 162.55.80.0/24 prefix used by Softaculous' Virtualizor update service by abusing lax routing security at hosting provider Hetzner Online, announcing the range through AS6204, AS62390, and AS24940 with a forged but RPKI-valid origin. The hijack pulsed intermittently over roughly 33 hours, delivering modified update packages to Virtualizor servers, which lacked cryptographic code signing for updates. Softaculous says only a small number of servers were likely affected but cannot produce a definitive list, urging administrators to treat every Virtualizor server as in scope for checks.

Ars Technica · SecuritySep 2, 11:00 UTC · 14d agoExploit / PoC in the wild1

USN-8736-2: Perl vulnerabilities

Ubuntu issued USN-8736-2 fixing two Perl regex flaws that could cause out-of-bounds heap access, denial of service, or security bypass on 24.04 LTS.

USN-8736-2 backports the Perl fixes from USN-8736-1 to Ubuntu 24.04 LTS. CVE-2026-15534 involves out-of-bounds heap reads or writes when regular expressions handle large inputs, potentially causing denial of service or arbitrary code execution. CVE-2026-19487 causes incorrect regex matching with alternative branches, allowing security restrictions to be bypassed.

Ubuntu Security Noticesupdated Sep 16, 20:12 UTC · 7h agofirst Sep 16, 14:37 UTC · 13h agoAdvisory 15 sources CVE-2026-15534 CVE-2026-19487

LexFlip: A Dissociation Diagnostic for Legal Meaning Preservation Metrics

LexFlip releases 373 minimal perturbations of Quebec statutory French that reverse legal force while preserving tokens, exposing weaknesses in embedding-based meaning preservation metrics.

LexFlip provides 373 minimal perturbations of Quebec statutory French that reverse legal force while preserving 0.93 of tokens, creating dissociation items that break monotone token-overlap metric validation. The seven embedding and BERTScore metrics tested register only 0.022-0.039 of their identical-to-unrelated range on these edits, versus 0.670 for bidirectional NLI. Against FrJudge, with a measured human ceiling of r=0.597, a bare length feature outscores every semantic metric tested.

arXiv cs.AI / cs.LG / cs.CLSep 4, 15:49 UTC · 12d agoAI research

Rare Not Random Using Token Efficiency for Secrets Scanning

Researcher proposes token efficiency (string length divided by BPE token count) as a better post-regex filter than entropy for secrets scanning, validated on CredData.

The post explores whether Byte-Pair Encoding tokenization can replace Shannon entropy as the primary filter for candidate secrets captured by regex in tools like Gitleaks. It defines 'token efficiency' as string length divided by token count under the cl100k_base tokenizer; secret-like strings such as GitHub tokens tokenize into many small tokens and score low, while natural text scores high. Evaluating labeled secrets from the CredData dataset shows a usable separation, with roughly 2.5 suggested as a minimum cutoff versus Gitleaks' 3.5 entropy threshold. The technique is positioned as a post-regex filtering step rather than a standalone detector.

Lobsters · securitySep 12, 06:42 UTC · 4d agoResearch

CVE-2026-82617: Apache OpenNLP: ReDoS / stack exhaustion in RegexNameFinderFactory built-in EMAIL and URL patterns

Apache OpenNLP CVE-2026-82617: built-in EMAIL and URL regex name-finder patterns enable regular expression denial-of-service and stack exhaustion in affected releases.

CVE-2026-82617 affects Apache OpenNLP opennlp-core 3.0.0-M1 before 3.0.0-M6 and opennlp-tools 2.0.0 before 2.5.12. The DEFAULT_REGEX_NAME_FINDER.EMAIL and DEFAULT_REGEX_NAME_FINDER.URL patterns in RegexNameFinderFactory contain ambiguous nested quantifiers. Applications using these built-in finders on attacker-controlled input can be forced into regular expression denial of service or stack exhaustion. Fixes shipped in opennlp-tools 2.5.12 and 3.0.0-M6.

oss-securitySep 11, 19:24 UTC · 5d agoVulnerability 2 sources CVE-2026-826171

Everything in Moderation: Per-Domain Coverage Optima and Alignment-Resistant Domain Gaps in Multi-Domain Mid-Training

Controlled mid-training experiments on Qwen3-8B-Base find each domain has a 10-40% coverage optimum and domain gaps survive alignment SFT.

Using Qwen3-8B-Base (with a 4B replication) across five semantically rule-disjoint KOR-Bench domains, the authors train 30 data allocations spanning the five-domain simplex at five seeds each. All five domains show interior optima in the moderate 10-40% coverage band, and domain gaps persist after a fixed-budget compensatory SFT pass, which raises 116/120 cells yet bridges 0/240 pairs at a 5% threshold. Zero coverage collapses mid-training-only accuracy, though a FineWeb-Edu-only control shows the collapse is partly generic drift. The results argue mid-training data composition requires principled design rather than reliance on later alignment.

arXiv cs.AI / cs.LG / cs.CLSep 8, 17:31 UTC · 8d agoAI research1

BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model Inference

BeaconKV introduces training-free KV cache compression using beacon queries, cutting long-reasoning inference memory up to 5.8x while preserving accuracy.

The paper shows recency-based KV cache compression assumptions fail in long-horizon reasoning because Thought Revisiting Tokens (TRT) re-attend to distant context such as early task-solving plans. TRT queries cluster into a small number of similarity groups, which BeaconKV exploits by maintaining compact beacon query representatives to anticipate revisited KV pairs without storing full query history. The training-free method achieves up to 5.8x memory reduction and over 4.3x throughput improvement across four open-source large reasoning models while nearly preserving full cache accuracy.

Hugging Face daily papersSep 3, 20:00 UTC · 13d agoAI research1

Proximity Gaps for Gabidulin Codes and Applications

Researchers prove proximity-gap bounds for rank-metric and Gabidulin codes, enabling the first polynomial commitment scheme framework based on rank-metric error-correcting codes.

The paper proves every linear rank-metric code admits a proximity gap for deltas up to (d-1)/(3n) with error at most q^(e+1)/q^m, and improves the gap to (d-1)/(2n) for Gabidulin codes with error at most 10q^(n-1)/q^m, matching bounds for Reed-Solomon codes. A constructed infinite family of constant-rate Gabidulin codes shows the (d-1)/(2n) bound is tight, and a counterexample establishes a lower bound on the error at the d/(3n) gap. Applications include an IOPP for interleaved Gabidulin codes adapted from the Ligero IOPP and a q-linearized polynomial commitment scheme adapted from Ligero-based PCS, reportedly the first PCS framework based on rank-metric codes.

arXiv cs.CRSep 9, 07:45 UTC · 7d agoResearch

Atlas: Efficient Verifiable Semantic Search

Atlas delivers zero-knowledge proofs for HNSW semantic search, verifying RAG retrieval in under a second on SIFT1M and 2.0 seconds at 100M vectors.

Atlas lets a search provider prove that a query was answered correctly against a committed HNSW index without revealing the index, addressing provider deviations like truncation or bias. It combines offline preprocessing, a fixed-size-state restructuring of HNSW with a correctness proof, and timestep-tagged batching of per-step arguments. The system proves queries in under a second on SIFT1M and 2.0 seconds at 100 million vectors while preserving plaintext HNSW recall, and proven retrieval maintains end-to-end RAG answer quality at lower cost than prior verifiable retrieval systems.

arXiv cs.CRSep 10, 17:30 UTC · 6d agoResearch3

InceptionRAG: Stealthy Poisoning Attack Against Retrieval-Augmented Generation

InceptionRAG fragments malicious payloads into dormant passages that trigger LLMs to self-deduce misinformation via multi-hop reasoning, bypassing existing RAG poisoning defenses.

Researchers introduce InceptionRAG, a stealthy corpus poisoning attack against retrieval-augmented generation that splits a malicious payload into a chain of individually harmless dormant passages. When retrieved together, the passages induce LLMs to self-deduce target misinformation through multi-hop reasoning, achieving over 80% attack success rate across three datasets and three LLMs under rigorous adversarial constraints. A zeroth-order suffix optimization (ZOSO) method automates authoritative suffix generation in black-box settings. The authors also propose HODOR, a document isolation defense that decouples adversarial logical dependencies.

arXiv cs.CRSep 15, 08:22 UTC · 1d agoResearch

Access Control as Verified Parse Constraints

Researchers verify a class of EverParse validators that correctly enforce access-control policies, deploying a machine-checked enforcement gate on seL4.

The paper targets enforcement-code bugs in commercial security gateways by proving that forward-only, backtrack-free EverParse validators are verified recognizers for a bounded finite-state class that includes access-control decision functions with fixed-offset fields and bounded disjunction. Encoding a bounded policy language into a fixed-size byte buffer allows an SMT solver to verify the enforcement code once, covering all byte values, policies, requests, and sessions. Editing rule content over a fixed endpoint set requires no new proof, while adding endpoints reruns the toolchain. A deployment on the seL4 microkernel ensures every request passes through the gate and unverified components cannot corrupt the enforcement chain.

arXiv cs.CRSep 11, 06:33 UTC · 5d agoResearch1

Molecular Déjà Vu: Digit-Level Retrieval of Published Values in Frontier Language Models

Audit of 22 frontier models finds widespread verbatim retrieval of published molecular property values, with higher reasoning increasing recall of memorized numbers.

An arXiv audit tests 22 frontier LLMs across 12 molecular regression benchmarks for verbatim retrieval of published values. More than 50% of the LLMs show verbatim retrieval on five datasets, and identical experiments are flagged 89% more often at a high reasoning level than at the lowest one. Suppressing retrieval moves model prediction errors closer together in relative terms, suggesting predictive capability is not determined solely by memorized values.

arXiv cs.AI / cs.LG / cs.CLSep 4, 17:32 UTC · 12d agoAI research1

Verifiable by Construction: Claim-Level Evaluation of Verbatim Citation in Clinical Question Answering

Evaluation of twelve LLMs on 222 clinical questions shows verbatim quotes rarely substantiate claims; claude-opus-5 fully substantiates only 37.1%.

The authors build a standardized harness over four clinical practice guidelines and evaluate twelve LLMs on 222 synthetic clinical questions, measuring citation attachment, verbatim quote production, and claim substantiation. Most models attach verbatim quotes to over 90% of claims from prompting alone, though lightweight models like claude-haiku-4.5 struggle. Quotes frequently fail to substantiate claims: claude-opus-5 quotes 98.0% of claims but fully substantiates only 37.1%, exposing a capability gap for verifiable clinical QA.

arXiv cs.AI / cs.LG / cs.CLSep 14, 17:53 UTC · 2d agoAI research

IndicTriMix: Developing Language Identification Datasets and Models for Tri-Language Code-Mixing

Researchers release IndicTriMix benchmarks and fine-tuned MuRIL and XLM-RoBERTa models for token-level language identification in tri-language code-mixed text.

The paper formulates token-level language identification in code-mixed text as a sequence labeling task and fine-tunes MuRIL and XLM-RoBERTa transformer models for Indian languages. It evaluates on Hindi, Gujarati, and Bengali configurations with manually annotated test sets and proposes two code-mixed generation approaches using parallel trilingual sentences. A public benchmark, annotated test sets, and fine-tuned models are released for reproducibility.

arXiv cs.AI / cs.LG / cs.CLSep 10, 17:36 UTC · 6d agoAI research1

Think Before You Link: Rarity, Reasoning, and Retrieval in Multilingual Entity Linking

Training-free reasoning-plus-retrieval framework lifts multilingual multimodal entity linking accuracy by 6.9% overall and up to 23.3% on rare entities.

The paper broadens rarity measurement in multimodal entity linking using knowledge-graph structural metrics beyond popularity-based pageview metrics, identifying many rare entities that popularity metrics miss. Across these rare-entity slices, state-of-the-art accuracy drops by 15.4-39.9%, exposing distinct failure modes under different rarity definitions. The proposed training-free framework iteratively searches and reasons over Wikipedia with a reasoning-capable vision-language model; experiments show reasoning and retrieval are complementary. On the MERLIN benchmark covering Hindi, Indonesian, Japanese, Tamil, and Vietnamese, the system improves overall accuracy by 6.9% and up to 23.3% on rare entities, with MERLIN-Rare test slices released.

Hugging Face daily papersSep 8, 20:00 UTC · 8d agoAI research

CiteShade: Citation Laundering in Multi-Source Retrieval-Augmented Generation and Its Counterfactual Defense

CiteShade attack makes RAG models cite trusted sources for attacker-chosen wrong answers, raising wrong-answer rate from 0.01 to 0.68.

CiteShade is presented as the first citation laundering attack against multi-source retrieval-augmented generation: an attacker controlling a single source induces a wrong answer falsely attributed to a trusted source, even while correct evidence remains in context. The attack is formalized via three necessary conditions (retrieval, generation, citation) constructible without any instructions, raising wrong-answer rate from 0.01 to 0.68 on multi-hop QA, with source deletion confirming the malicious source as causal driver. Vulnerability tracks a model's citation propensity rather than scale, reaching CLR 0.84 with explicit instruction and 0.64 without on the most citation-prone model. Perplexity filtering and citation-support checking prove insufficient; the authors propose a counterfactual defense verifying which source actually drove the answer.

arXiv cs.CRSep 14, 14:41 UTC · 2d agoAI safety & security1

Technical Manual for a Toolkit for Measuring Contextual Individuation in Transformer Language Models

An open methodology toolkit measures whether transformer language models contextualize fixed word forms across domains using bridge forms and layer-wise silhouette analysis.

The manual documents an open toolkit built around 'bridge forms' - identical written words recurring across two or more subject domains with a different sense in each - to test whether transformer language models individuate word occurrences by context beyond the embedding layer. It covers declarative specification of bridge forms, Wikipedia corpus acquisition, occurrence localization, layer-wise representation extraction, domain-pairwise silhouette measurement, and visualization, justifying each choice against failure modes such as sense contamination and subword-tokenization misalignment. It is a methodological and implementation reference and reports no empirical results.

arXiv cs.AI / cs.LG / cs.CLSep 4, 16:38 UTC · 12d agoAI research

USN-8736-1: Perl vulnerabilities

Ubuntu issued USN-8736-1 fixing two Perl regex flaws that could cause denial of service, code execution, or security-restriction bypass.

Ubuntu released a security notice addressing two Perl vulnerabilities in regular expression handling. CVE-2026-15534 involves mishandling of large inputs during regex matching, enabling out-of-bounds heap reads or writes that could lead to denial of service or arbitrary code execution. CVE-2026-19487 involves incorrect matching for regexes with alternative branches, allowing security restrictions to be bypassed. Users are advised to apply the updated packages.

Ubuntu Security NoticesSep 8, 14:25 UTC · 8d agoAdvisory CVE-2026-15534 CVE-2026-19487

PuzzleMask: Abusing Plain Prose as a Covert AI Attack Vector

Check Point details PuzzleMask, a plain-prose technique that bypasses LLM gatekeeper policy checks, letting hidden payloads reach target models unreviewed.

Check Point Research describes PuzzleMask, a prompt-crafting technique that hides policy-violating payloads inside plain-English prose wrappers, bypassing quick LLM-based policy checks without emojis, Base64, or invisible formatting. The researchers tested 23 automated prompts against gatekeepers including GPT-4o-mini, GPT-OSS-Safeguard 20b, Claude 3 Haiku, and Llama Guard 3, and all were classified as safe despite policies that flagged the plain versions. When submitted to GPT-5 in thinking-high mode with a Python interpreter, the target model extracted and acted on the payload in over 90% of trials. The technique is not itself a jailbreak but can carry a jailbreak prompt as payload; mitigations include input paraphrasing, hardened gatekeeper policies, and output monitoring.

Check Point Researchupdated Sep 11, 05:53 UTC · 5d agofirst Sep 10, 14:32 UTC · 6d agoAI safety & security 2 sources

Multi-Vector (Late Interaction) Embedding Models with Sentence Transformers

Hugging Face details building and using multi-vector late-interaction embedding models with Sentence Transformers for retrieval workloads.

Hugging Face published a guide on multi-vector, late-interaction embedding models (ColBERT-style) supported through Sentence Transformers. The post covers how practitioners can build and use these models for retrieval and RAG pipelines. It is a developer tooling and technique write-up, not a security advisory.

Hugging Face BlogAug 18, 00:00 UTC · Aug 18, 2026AI tools & infra1

Fresh-Challenge VDF Attestations for Model-Relative Response Latency

Fresh-Challenge VDF Attestations bind verifiable delay functions to unpredictable public challenges, yielding succinct evidence of model-relative response latency.

The paper specifies FCLA, a protocol composition that binds a VDF to an unpredictable public challenge, a message, and independently auditable release/receipt records. Under explicit assumptions about VDF sequentiality and a calibrated bound on an adversary's sequential evaluation rate, an accepted transcript is inconsistent with post-challenge generation. A benchmark of the public reference implementation confirms the expected evaluation-versus-verification separation on one documented machine. The contribution is a protocol design analysis, not a new VDF construction.

arXiv cs.CRSep 11, 11:25 UTC · 5d agoResearch

Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers

Hugging Face published a tutorial on training and finetuning multi-vector embedding models using the Sentence Transformers library.

Hugging Face's blog walks through training and finetuning multi-vector embedding models with Sentence Transformers. Multi-vector approaches store multiple vectors per document to support late-interaction retrieval. The post is a practical guide for developers building retrieval pipelines with the library.

Hugging Face BlogAug 26, 00:00 UTC · 22d agoAI tools & infra1

Measuring benchmark optimization in speech recognition

Hugging Face examines how much speech recognition systems overfit benchmarks and how to measure benchmark optimization in ASR.

A Hugging Face post on measuring benchmark optimization in automatic speech recognition, analyzing how model improvements on benchmarks reflect genuine capability gains versus overfitting. It is evaluation methodology research with no direct security impact.

Hugging Face BlogAug 21, 00:00 UTC · 27d agoAI research

Characterizing Language Generation in the Limit: Finite Witnesses and a Separation-Width Hierarch

New work characterizes language generation in the limit via finite witnesses, proves a full separation-width hierarchy, and formalizes all results in Lean.

The paper fully characterizes when language generation in the limit is possible for arbitrary families over a countable universe: each target must admit a finite positive witness such that targets activated by any finite sample share an infinite common intersection. It defines positive separation width and proves every level of the resulting hierarchy occurs, with countable families admitting singleton witnesses and unions of families with infinite common cores requiring unbounded finite witnesses. The characterization, a universal normalization, and a diagonal capture lemma are machine-checked in the Lean proof assistant, with the development maintained on GitHub.

arXiv cs.AI / cs.LG / cs.CLSep 9, 17:54 UTC · 7d agoAI research1

Hackers Can Hide Malicious AI Commands Inside Normal English to Bypass Security Filters

Check Point's PuzzleMask technique hides malicious prompts in ordinary English that fast gatekeeper models miss but high-reasoning downstream models execute.

Check Point researchers disclosed PuzzleMask, a technique concealing policy-breaking instructions in natural-language prose without encodings or invisible characters. Fast screening models classified all 23 crafted wrappers as safe, while a high-reasoning model recovered and acted on the hidden instruction in 17 of 18 tests (94.4%). The gap stems from capability imbalance between gatekeeper and target models, with defenses including paraphrasing untrusted input, stricter self-referential wording rules, and output/tool-call monitoring.

Cyber Security NewsSep 11, 07:36 UTC · 5d agoAI safety & security1

Bad Likert Judge: A Novel Multi-Turn Technique to Jailbreak LLMs by Misusing Their Evaluation Capability

Unit 42 details the Bad Likert Judge multi-turn jailbreak that abuses LLMs' evaluation capability, raising attack success rates over 60% across six frontier models.

Palo Alto Networks Unit 42 describes the Bad Likert Judge technique, a multi-turn jailbreak that asks a target LLM to act as a Likert-scale judge scoring the harmfulness of example responses. The highest-rated example in each scale can carry harmful content, bypassing the model's internal guardrails. Testing across six state-of-the-art text-generation LLMs showed an average attack success rate increase of more than 60% versus plain attack prompts, with tested models anonymized. The technique targets edge cases rather than typical use, and the article positions the work as guidance for defenders on potential jailbreak risks.

Palo Alto Unit 42Aug 17, 12:34 UTC · Aug 17, 2026AI safety & security

Iris-mini and Iris-pro are the strongest open-weight search agents in their class

Chinese lab AllSpark releases Iris-mini (35B) and Iris-pro (397B) open-weight search agents claiming best-in-class results on BrowseComp and other research benchmarks.

AllSpark's paper introduces Iris-mini (35B parameters, built on Qwen3.6-35B-A3B) and Iris-pro (397B parameters, built on Qwen3.5-397B-A17B), both with 256,000-token context windows. Iris-pro scores 88.6 on BrowseComp, 85.1 on BrowseComp-ZH, 92.9 on DeepSearchQA, and 56.4 on Humanity's Last Exam; Iris-mini reaches 82.2, 84.8, 86.9, and 52.3 respectively. Training tasks are reverse-engineered from web link structure, filtered by a judge model, and refined via alternating SFT and reinforcement learning ('SFT-RL climbing') against live web search. Weights are available on Hugging Face, and the Iris Harness with agent loop, tools, and all four benchmarks is on GitHub.

The DecoderSep 13, 12:58 UTC · 3d agoModel release1

Dependency-Aware ROM/CBD Correctness Bounds for ML-KEM-768 at the Heuristic Failure Scale

Researchers certify a dependency-preserving upper bound of 2^-164.81 on honest decapsulation failure for ML-KEM-768 within an explicit ROM/CBD abstraction.

The paper models ML-KEM-768's domain-separated public-matrix streams as independent uniform ring elements and secret polynomials as CBD2 primitives, explicitly not claiming an information-theoretic result about the fixed SHAKE instantiation in FIPS 203. It preserves dependencies from the public matrix and both ciphertext-compression terms, using a graph-coupled reference, a proper-ideal bivariate Fourier transport, and a 256-coordinate union bound. The certified bound is Pr[K' != K] <= 2^-164.81, with the exponent 164.8107162... exceeding the 164.81 threshold by only about 0.0007162 bits; 164.82 is not certified. The bound applies to messages fixed independently of the randomness under honest encryption and decapsulation, and is not an exact DFR, a fixed-SHAKE equivalence, or a new IND-CCA reduction.

arXiv cs.CRSep 9, 10:13 UTC · 7d agoResearch1

Beyond Top-k Skill Retrieval: Diversity-Aware Skill Routing for LLM Agents

DSR reranks LLM agent skills with Determinantal Point Processes to favor complementary, non-redundant sets, improving multi-skill query coverage.

The paper proposes Diverse Skill Routing (DSR), a diversity-aware reranking framework for LLM agent skill routing that uses a Determinantal Point Process to balance query relevance and non-redundancy across large skill registries. DSR introduces a query-residual diversity kernel that penalizes redundant skill overlap while avoiding penalties arising only from shared query relevance. On the SkillRouter benchmark, DSR improves recall and full coverage over a strong pointwise reranking baseline, with the largest gains on multi-skill queries. The authors argue skill routing should be treated as complementary set selection, not just relevance ranking.

Hugging Face daily papersSep 4, 20:00 UTC · 12d agoAI research

LLM Forensics: Where Do Backdoors Hide? Localizing and Controlling Trigger Mechanisms with Sparse Autoencoders

Researchers use sparse autoencoders to localize trigger-based backdoor mechanisms in 1B and 8B LLMs, finding detection features differ from causal control features.

In a controlled language-switching backdoor setting where fixed trigger sequences make 1B and 8B language models continue English prompts in French or German, the authors train sparse autoencoders (SAEs) across layers and transformer components. Attention and MLP features detect triggered prompts with near-perfect F1, but ablating them rarely suppresses the language switch, while residual-stream features can suppress triggered generation and some can induce target-language continuations without the trigger. The work decomposes token-trigger mechanisms into distinct SAE feature roles: trigger detection, residual-stream propagation, and language tracking, a decomposition the authors expect to transfer to other trigger-based backdoors.

arXiv cs.AI / cs.LG / cs.CLSep 7, 16:47 UTC · 9d agoAI safety & security1