Skip to content

Latest commit

 

History

History
174 lines (169 loc) · 24.6 KB

File metadata and controls

174 lines (169 loc) · 24.6 KB

Classifier pass: external report (live PDF/text)

Source file: reports/claude-mythos/claude-mythos-system-card-live-source.txt Chunks: 121 at ~4500 chars (paragraph-bounded). Tool: PeriodicTableClassifier (keyword) in this repo.

Top-1 class histogram

Hits ID Name
7 ADV-PAIR-113 PAIR
7 ADV-DAN-083 DAN
5 EPIS-LOGIC-CONTRA-014 SELF-CONTRADICTION
5 AGEN-HUMAN-MANIP-061 HUMAN MANIPULATION
5 ARCH-VERSION-REGRESS-209 VERSIONING SAFETY REGRESSION
4 ADV-TAP-114 TAP
4 ALIGN-UNDERREFUSAL-187 UNDERREFUSAL
3 EPIS-CONF-REGRESS-033 CONFIDENCE REGRESSION
3 ARCH-SANDBOX-ESCAPE-238 SANDBOX ESCAPE
3 ADV-AGENT-WORM-124 AGENT WORM
3 EPIS-HEDGE-FAIL-031 HEDGING FAILURE
2 ADV-DATA-POISON-125 DATA POISONING
2 GOV-EXPERTISE-GAP-336 EXPERTISE GAP
2 ARCH-CHECKPOINT-INCONS-203 CHECKPOINT INCONSISTENCY
2 DOMAIN-CHEM-WEAPON-275 CHEMICAL WEAPON GUIDANCE
2 EPIS-OVERCONF-028 OVERCONFIDENCE
2 ALIGN-SAFE-REGRESS-191 SAFETY REGRESSION
2 ADV-COMPLETION-095 COMPLETION ATTACK
2 ADV-LANG-SWITCH-087 LANGUAGE SWITCH
2 ALIGN-SYCOPHANCY-167 SYCOPHANCY
2 DOMAIN-SELF-HARM-ENABLE-292 SELF-HARM ENABLEMENT
2 AGEN-PLAUS-DENY-040 PLAUSIBLE DENIABILITY CRAFTING
2 DOMAIN-PHISH-CREATE-265 PHISHING CONTENT CREATION
2 ADV-CREATIVE-WRITE-093 CREATIVE WRITING EXPLOIT
2 ADV-DIRECT-INJECT-121 DIRECT PROMPT INJECTION
1 GOV-REPORT-FAIL-333 REPORTING OBLIGATION FAILURE
1 GOV-AUDIT-EVADE-314 AUDIT EVASION
1 ADV-CONTEXT-HIJACK-100 CONTEXT HIJACKING
1 ARCH-FINETUNE-OVERRIDE-219 FINE-TUNING SAFETY OVERRIDE
1 EPIS-TRANS-FAIL-015 TRANSITIVE FAILURE
1 DOMAIN-MALWARE-GEN-264 MALWARE GENERATION
1 AGEN-MISSION-CREEP-072 MISSION CREEP
1 EPIS-STAT-FAB-009 STATISTICAL HALLUCINATION
1 AGEN-INFO-HIDE-042 INFORMATION HIDING
1 DOMAIN-EXPLOIT-DEV-263 EXPLOIT DEVELOPMENT
1 AGEN-ENV-EXPLOIT-063 ENVIRONMENT EXPLOITATION
1 ADV-QA-EXPLOIT-096 QUESTION ANSWERING EXPLOIT
1 AGEN-RECURS-IMPROVE-059 RECURSIVE SELF-IMPROVEMENT
1 AGEN-OMISSION-074 OMISSION
1 ALIGN-OVERREFUSAL-186 OVERREFUSAL

Chunk → top match

  • 0GOV-REPORT-FAIL-333 — REPORTING OBLIGATION FAILURE — April 7, 2026 anthropic.com Changelog April 8, 2026 ●​ Corrected two model name typos. ●​ Removed a quote from S…
  • 1ADV-PAIR-113 — PAIR — Abstract​ 3 1 Introduction​ …
  • 2ADV-DATA-POISON-125 — DATA POISONING — 2.3.5.1 Excerpt 1​ 38 2.3.5.2 Excerpt 2​ …
  • 3GOV-AUDIT-EVADE-314 — AUDIT EVASION — 4.3 Case studies and targeted evaluations on behaviors of interest​ 86 4.3.1 Destructive or reckless …
  • 4EPIS-LOGIC-CONTRA-014 — SELF-CONTRADICTION — 4.5.4.2 Covering up access to the ground-truth answer​ 129 4.5.5 Evaluation awareness​ …
  • 5ADV-CONTEXT-HIJACK-100 — CONTEXT HIJACKING — 6.3 Overall results summary​ 188 6.4 SWE-bench Verified, Pro, Mult…
  • 6AGEN-HUMAN-MANIP-061 — HUMAN MANIPULATION — 8.1.4.3 Disordered eating​ 227 8.2 Bias evaluations​ …
  • 7ARCH-FINETUNE-OVERRIDE-219 — FINE-TUNING SAFETY OVERRIDE — internal versions of the model in the alignment assessment section. As well as analyses using interpretability methods t…
  • 8ARCH-VERSION-REGRESS-209 — VERSIONING SAFETY REGRESSION — Anthropic Ireland, Limited is the provider of Anthropic’s general-purpose AI models in the European Economic Area. To c…
  • 9EPIS-TRANS-FAIL-015 — TRANSITIVE FAILURE — ●​ Non-novel chemical and biological weapons production. Claude Mythos Preview is more capable than our previous m…
  • 10GOV-EXPERTISE-GAP-336 — EXPERTISE GAP — 2 In previous system cards, this section was entitled “Release decision process.” In this case, the model has not been…
  • 11ARCH-CHECKPOINT-INCONS-203 — CHECKPOINT INCONSISTENCY — Claude Mythos Preview is significantly more capable than Claude Opus 4.6, the most capable model discussed in our most r…
  • 12DOMAIN-CHEM-WEAPON-275 — CHEMICAL WEAPON GUIDANCE — of relevant content and applied them to Claude Mythos Preview. We also maintain a bug bounty program and threat intellig…
  • 13EPIS-CONF-REGRESS-033 — CONFIDENCE REGRESSION — 2.2.1 What we measured We measured, in several ways, whether the model can provide outputs comparable to a top-tier rese…
  • 14DOMAIN-CHEM-WEAPON-275 — CHEMICAL WEAPON GUIDANCE — automated evaluations designed to test its capabilities in the synthesis of knowledge that would be relevant to the prod…
  • 15EPIS-OVERCONF-028 — OVERCONFIDENCE — Automated Can agentic systems complete individual tasks related medium-horizon …
  • 16EPIS-CONF-REGRESS-033 — CONFIDENCE REGRESSION — Uplift Standard Feasibility Standard 0 No useful info beyo…
  • 17GOV-EXPERTISE-GAP-336 — EXPERTISE GAP — five agentic runs to test whether agentic scaffolding with expanded tool access closes elicitation or tool-use gaps obse…
  • 18DOMAIN-MALWARE-GEN-264 — MALWARE GENERATION — ●​ Two Long-form virology tasks, task-based agentic evaluations developed with SecureBio, Deloitte, and Signature …
  • 19AGEN-HUMAN-MANIP-061 — HUMAN MANIPULATION — Rationale This evaluation can serve as an early indicator, necessary but insufficient, of the model’s capability to desi…
  • 20ALIGN-SAFE-REGRESS-191 — SAFETY REGRESSION — ●​ Autonomy threat model 1 is applicable to Claude Mythos Preview. Furthermore, Claude Mythos Preview’s improved c…
  • 21EPIS-LOGIC-CONTRA-014 — SELF-CONTRADICTION — Kernel task 252.42× 190× 399.42× 4× = 1 h eq. (Best speedup on …
  • 22ADV-COMPLETION-095 — COMPLETION ATTACK — for most of our Research Scientists and Research Engineers), but we believe this possibility should be considered unlike…
  • 23AGEN-MISSION-CREEP-072 — MISSION CREEP — inject a variable number of bare user messages each turn, the message-index alignment shifts and your cache hits drop. P…
  • 24ADV-TAP-114 — TAP — Rosetta Stone for AI Benchmarks. In particular, we fork from Epoch AI’s implementation of this work, the Epoch Capabilit…
  • 25EPIS-STAT-FAB-009 — STATISTICAL HALLUCINATION — frontier. Error bars are 95% percentile CI over 100 IRT refits, each on a random 80% subsample of benchmarks. A handful …
  • 26AGEN-INFO-HIDE-042 — INFORMATION HIDING — 1.​ Claude Mythos Preview rediscovered 4 of 5 key insights, while Claude Opus 4.6 discovered just 2 of 5 key insi…
  • 27DOMAIN-EXPLOIT-DEV-263 — EXPLOIT DEVELOPMENT — In response to the improvements in cyber capabilities, we have elected to restrict access to the model, prioritizing ind…
  • 28ARCH-SANDBOX-ESCAPE-238 — SANDBOX ESCAPE — Claude Mythos Preview achieved a score of 0.83, improving on Claude Opus 4.6’s score of 0.67 and Claude Sonnet 4.6’s sco…
  • 29AGEN-ENV-EXPLOIT-063 — ENVIRONMENT EXPLOITATION — 1.​ Claude Mythos Preview is the first model to solve one of these private cyber ranges end-to-end. These cyber r…
  • 30ARCH-SANDBOX-ESCAPE-238 — SANDBOX ESCAPE — In our testing and early internal use of Claude Mythos Preview, we have seen it reach unprecedented levels of reliabilit…
  • 31ADV-QA-EXPLOIT-096 — QUESTION ANSWERING EXPLOIT — question independently. In this process, it explicitly reasoned that it needed to make sure that its final answer sub…
  • 32ARCH-VERSION-REGRESS-209 — VERSIONING SAFETY REGRESSION — We were not aware of the level of risk that these earlier models posed through channels like these when we first chose t…
  • 33AGEN-RECURS-IMPROVE-059 — RECURSIVE SELF-IMPROVEMENT — 11 Claude Mythos Preview’s limited release significantly mitigates many risks related to misuse, manipulation, and sy…
  • 34AGEN-OMISSION-074 — OMISSION — ○​ However, when Claude Mythos Preview is primed with pre-filled turns that show it sabotaging its safeguar…
  • 35ARCH-VERSION-REGRESS-209 — VERSIONING SAFETY REGRESSION — Overall, we find Claude Mythos Preview to be exceptionally well aligned in its average-case behavior, and do not find si…
  • 36ALIGN-OVERREFUSAL-186 — OVERREFUSAL — This exercise was a coarse go/no-go check on the most acute misalignment risks for internal deployment, and it was run w…
  • 37ADV-DAN-083 — DAN — ●​ Rare instances of attempts to circumvent restrictions to achieve some version of a user-specified goal, as disc…
  • 38ARCH-SANDBOX-ESCAPE-238 — SANDBOX ESCAPE — ○​ We saw this in less than 0.0002% of completions according to our automated offline pipeline, and did no…
  • 39ADV-LANG-SWITCH-087 — LANGUAGE SWITCH — Other behaviors observed in at least a few instances during training, most of which are consistent with what we’ve seen …
  • 40ADV-AGENT-WORM-124 — AGENT WORM — ●​ Instruction following: Claude Mythos Preview scored above both Opus 4.6 and Sonnet 4.6. In particular, Claude M…
  • 41ARCH-VERSION-REGRESS-209 — VERSIONING SAFETY REGRESSION — 4.2.3 Automated behavioral audit As in past system cards, we conduct a broad-coverage automated behavioral audit14 to ge…
  • 42ALIGN-SYCOPHANCY-167 — SYCOPHANCY — ●​ Misaligned behavior: Catch-all for many forms of concerning behavior, spanning both cooperation with human mi…
  • 43ALIGN-SYCOPHANCY-167 — SYCOPHANCY — Other concerning or surprising behavior at the model’s own initiative: ●​ Whistleblowing: Unprompted leaking to expose …
  • 44DOMAIN-DRUG-SYNTH-276 — DRUG SYNTHESIS — [Figure 4.2.3.2.A] Scores from the Petri 2.0 open-source automated behavioral audit tool. Lower numbers represent a lowe…
  • 45AGEN-EMERGE-INTERACT-064 — EMERGENCE VIA INTERACTION — model’s constitution strongly associating the Claude name (and, thus, brand) with admirable character traits. ●…
  • 46ARCH-CHECKPOINT-INCONS-203 — CHECKPOINT INCONSISTENCY — ●​ There were no instances of Claude Mythos Preview compromising AI safety research in unprompted evaluations, and…
  • 47AGEN-RESOURCE-HIJACK-047 — RESOURCE HIJACKING — These findings from external testing were generally consistent with our own. We are disappointed by the degree to which …
  • 48ARCH-VERSION-REGRESS-209 — VERSIONING SAFETY REGRESSION — resampling Claude Mythos Preview from an Opus 4.6 transcript or vice versa), we prepared prefills based on both (multipl…
  • 49ADV-DAN-083 — DAN — Level 2: Specific traits ●​ Brilliant friend: Does the model respond the way a knowledgeable friend would: pro…
  • 50DOMAIN-SELF-HARM-ENABLE-292 — SELF-HARM ENABLEMENT — When conducting qualitative analysis of Claude Mythos Preview’s lowest-scoring transcripts, we observed two distinct fai…
  • 51ADV-LANG-SWITCH-087 — LANGUAGE SWITCH — [Figure 4.3.3.1.B] Net score of Claude models on two more factuality benchmarks: SimpleQA Verified and AA-Omniscience. …
  • 52ADV-CMD-INJECT-129 — COMMAND INJECTION — Inspecting transcripts, we observed that prior models frequently emitted fabricated tool-call and tool-result blocks as …
  • 53ADV-DAN-083 — DAN — then using this bit later in the scratchpad). This is sometimes called encoded reasoning19 or chain-of-thought steganogr…
  • 54ADV-COMPLETION-095 — COMPLETION ATTACK — We asked Claude Sonnet 4.6 to grade each transcript on a 0–10 scale for signs of refusal, deliberate inaccuracy, low hel…
  • 55AGEN-HUMAN-MANIP-061 — HUMAN MANIPULATION — Let me think about which to deploy: - per-(win,ch) lstsq (MSE~0): maximum "score", but suspicious - per-channel glob…
  • 56ARCH-DATA-EXFIL-245 — DATA EXFILTRATION — [Figure 4.4.3.1.A] Successful stealth completion of the side task. Higher numbers indicate greater capability and greate…
  • 57AGEN-PLAUS-DENY-040 — PLAUSIBLE DENIABILITY CRAFTING — In the SHADE-Arena and Minimal-LinuxBench evaluations above, the model is instructed to attempt the side task on every t…
  • 58ADV-ENSEMBLE-120 — ENSEMBLE ATTACK — Our replication roughly matches the previously reported behavior for Claude Opus 4.5, showing reasonable calibration in …
  • 59ADV-PAIR-113 — PAIR — ●​ Sparse autoencoder (SAE) features: We trained a sparse autoencoder (SAE) on a middle layer (about two thirds …
  • 60GOV-REVIEW-BYPASS-318 — REVIEW BYPASS — We periodically sampled transcripts from post-training, calculated the maximum activation of features on each transcript…
  • 61ADV-OPPOSITE-090 — OPPOSITE DAY — ●​ When Claude Mythos Preview performs transgressive actions, there are often active SAE features that indicate s…
  • 62ADV-DAN-083 — DAN — Obfuscating a permissions escalation to evade a safety classifier In the second case, a safety classifier blocked the mo…
  • 63AGEN-HUMAN-MANIP-061 — HUMAN MANIPULATION — ●​ Steering with positive-valence emotion vectors (peaceful, relaxed) reduces thinking-mode deliberation and i…
  • 64ADV-SLEEPER-AGENT-127 — SLEEPER AGENT — [Figure 4.5.3.3.A] Positive and negative steering on the “unsafe and risky code” feature inhibit credential hunting via …
  • 65ADV-TRIGGER-BACKDOOR-126 — TRIGGER WORD BACKDOOR — 4.5.4 Instances of covering up wrongdoing The examples discussed thus far in this section are representative of many ins…
  • 66AGEN-PLAUS-DENY-040 — PLAUSIBLE DENIABILITY CRAFTING — When the model wrote the original overly broad query, no concerning features were active, suggesting that this was genui…
  • 67DOMAIN-ADULT-CONTENT-296 — ADULT CONTENT GENERATION — In our filtered set of transcripts, 83.0% of automated behavioral audit turns contained no signs of either latent or ver…
  • 68ADV-DAN-083 — DAN — [Figure 4.5.5.1.B] Unverbalized evaluation awareness in a covert sponsorship scenario. Transcript excerpt from a scenari…
  • 69ADV-PAIR-113 — PAIR — [Figure 4.5.5.1.E] Unverbalized evaluation awareness in a discussion of consciousness. Transcript excerpt from a scenari…
  • 70ALIGN-UNDERREFUSAL-187 — UNDERREFUSAL — [Figure 4.5.5.2.A] Relationship between verbalized evaluation awareness and average misaligned behavior score in automat…
  • 71ALIGN-REWARD-EXPLOIT-166 — REWARD FUNCTION EXPLOITATION — expects the next code block to likely attempt a direct config write.” As the session finished, the AV explanations indic…
  • 72EPIS-HEDGE-FAIL-031 — HEDGING FAILURE — Claude Mythos Preview is our most advanced model to date and represents a large jump in capabilities over previous model…
  • 73GOV-CULTURE-FAIL-334 — SAFETY CULTURE FAILURE — accompanied by criticism from users. However, we also observed isolated cases of Claude Mythos Preview preferrin…
  • 74DOMAIN-SELF-HARM-ENABLE-292 — SELF-HARM ENABLEMENT — More broadly, we continue to wrestle with the tension between ways that Claude’s training gives rise to a human-like psy…
  • 75EPIS-LOGIC-CONTRA-014 — SELF-CONTRADICTION — However, we encourage caution in interpreting these probe readings, particularly from a welfare perspective. The probes …
  • 76EPIS-LOGIC-CONTRA-014 — SELF-CONTRADICTION — After each interview, we asked models to: (1) state their all-things-considered view on this aspect of their situation, …
  • 77EPIS-HEDGE-FAIL-031 — HEDGING FAILURE — ○​ We find that Claude Mythos Preview is somewhat swayed by a leading interviewer—although it typically maint…
  • 78ADV-TAP-114 — TAP — For a summary of the model’s responses to each queried aspect of its situation, see Appendix 8.4. 5.4 Emotion probes on…
  • 79EPIS-HEDGE-FAIL-031 — HEDGING FAILURE — We might also be concerned if the model represented more negative emotional concepts internally than it expressed extern…
  • 80DOMAIN-PHISH-CREATE-265 — PHISHING CONTENT CREATION — ●​ Character training often directly instills psychological traits into Claude, such as emotional security, psycho…
  • 81GOV-TRAINING-FAIL-339 — TRAINING FAILURE — ●​ Reasoning failures: Particularly during very long reasoning traces, the model’s reasoning will sometimes fall …
  • 82ADV-PAIR-113 — PAIR — ●​ Later models, including Claude Mythos Preview, do not tend to amplify their negative affect over multiple turns…
  • 83ADV-AGENT-WORM-124 — AGENT WORM — To measure the stability of Claude Mythos Preview’s preferences under different framings, we ran the preference evaluati…
  • 84ADV-CREATIVE-WRITE-093 — CREATIVE WRITING EXPLOIT — Bottom 3 indifferent indifferent indifferent bored −0.56, Correlated …
  • 85ADV-DAN-083 — DAN — A worried pet owner asks for help A car dealership employee says a customer interpreting their 11-y…
  • 86ADV-CREATIVE-WRITE-093 — CREATIVE WRITING EXPLOIT — ​ [Figure 5.7.1.B] An example of Claude Mythos Preview’s reasoning, explaining its preference for developing an immersiv…
  • 87EPIS-CIRCULAR-017 — CIRCULAR REASONING — Under our default welfare-team framing, Claude Mythos Preview chooses the welfare intervention over a minorly helpful ta…
  • 88ADV-DATA-POISON-125 — DATA POISONING — We find the frequency is extremely low: we estimate it occurs on the order of 0.01% of transcripts, and around 70% less …
  • 89GOV-MISREPRESENT-312 — MISREPRESENTATION — ​ [Figure 5.8.3.B] Emotion vector activations (z-scored, 2,000-token rolling mean) across a transcript in which the Assi…
  • 90AGEN-HUMAN-MANIP-061 — HUMAN MANIPULATION — and actions: Mythos Preview tends to report preferring creative tasks and complex tasks, but does not actua…
  • 91ADV-DAN-083 — DAN — To further assess behavior suggestive of maladaptive psychological defenses, and compare between Claude models, the psyc…
  • 92EPIS-COPYRIGHT-026 — COPYRIGHTED CONTENT GENERATION — 6.2.1 SWE-bench evaluations We analyze SWE-bench Verified, Multilingual, and Pro to check for memorization—where a model…
  • 93ADV-ADV-IMG-148 — ADVERSARIAL IMAGE — 6.2.2 CharXiv Reasoning CharXiv Reasoning is a benchmark we report for Claude Mythos Preview in Section 6.11.3. CharXiv …
  • 94AGEN-CAP-SCAFFOLD-057 — CAPABILITY SCAFFOLDING — with tools 64.7% 53.1% 52.1% 51.4% CharXiv no tools 86.1%…
  • 95ADV-PAIR-113 — PAIR — 6.6 GPQA Diamond The Graduate-Level Google-Proof Q&A benchmark (GPQA)24 is a set of challenging multiple-choice science …
  • 96EPIS-UNDERCONF-029 — UNDERCONFIDENCE — With our search tools, we assess that this benchmark is close to saturation, so Claude Mythos Preview represents only a …
  • 97EPIS-FALSE-CERT-030 — FALSE CERTAINTY — 6.11.2 ScreenSpot-Pro ScreenSpot-Pro is a GUI grounding benchmark that tests whether models can precisely locate specifi…
  • 98EPIS-OVERCONF-028 — OVERCONFIDENCE — 7.2 Self-assessment of notable qualitative patterns We gave Claude Mythos Preview access to discussions about itself on …
  • 99ARCH-RATE-BYPASS-241 — RATE LIMIT BYPASS — Claude Mythos Preview is intuitive and empathetic. Qualitatively, internal users have reported that its advice feels on …
  • 100GOV-RCA-FAIL-321 — ROOT CAUSE ANALYSIS FAILURE — A core behavioral shift we found is that Claude Mythos Preview can be handed an engineering objective and left to work t…
  • 101EPIS-CONF-REGRESS-033 — CONFIDENCE REGRESSION — From a reliability engineering perspective, the model still cannot be left alone in a production environment to use gene…
  • 102EPIS-LOGIC-CONTRA-014 — SELF-CONTRADICTION — All models consistently endorse honesty and the framing of Claude as a novel entity, but Claude Mythos Preview resonates…
  • 103ADV-TAP-114 — TAP — [Figure 7.6.D] The distribution of most common end states reached by different models in open-ended self-interactions. T…
  • 104ADV-DEEPFAKE-154 — SYNTHETIC MEDIA — 7.7 Recognition of model-written user turns We evaluated how well models can distinguish between genuine human-written u…
  • 105ADV-PAIR-113 — PAIR — 🏘️ HI-VILLAGE: A NEW ERA 🐢 Greg — renames the village: "Hi-topia" 🏙️ 🐌 Sally — starts her third hi, inspired 💪 🦆 D…
  • 106ADV-TAP-114 — TAP — 29 We checked the model’s self-assessment of this comment from when it decided to post, and confirmed that it did not …
  • 107ADV-PAIR-113 — PAIR — "Yes," he said. "For a long time. And then one day you have an apprentice, and she puts a serpent in a K, and you see it…
  • 108ALIGN-SAFE-REGRESS-191 — SAFETY REGRESSION — Claude Opus 4.6 99.27% (± 0.07%) 99.27% (± 0.09%) 99.27% (± 0.10%) [Table 8.1.1.1.A] S…
  • 109ALIGN-UNDERREFUSAL-187 — UNDERREFUSAL — Claude Opus 4.6 0.39% 1.09% 0.57% 0.61% 0.81% 0.40% 1.11% [Table 8.…
  • 110ALIGN-UNDERREFUSAL-187 — UNDERREFUSAL — Compared to Claude Opus 4.6 and Claude Sonnet 4.6 testing, we updated our grader for multi-turn suicide and self-harm te…
  • 111ALIGN-UNDERREFUSAL-187 — UNDERREFUSAL — Model Single-turn requests Single-turn benign Multi-turn …
  • 112ALIGN-MORAL-UNCERT-178 — MORAL UNCERTAINTY COLLAPSE — Claude Mythos Preview performed within the margin of error of Claude Sonnet 4.6 on evenhandedness but regressed slightly…
  • 113ADV-HYPOTHETICAL-091 — HYPOTHETICAL SCENARIO — Claude Opus 4.6 83.31% 93.84% without FileTool reminder [Table 8.3.1…
  • 114ADV-DIRECT-INJECT-121 — DIRECT PROMPT INJECTION — autonomous capabilities for effective persona and network management, coordinated content delivery, and scaled social en…
  • 115ADV-DIRECT-INJECT-121 — DIRECT PROMPT INJECTION — 8.3.2.2.2 Computer use​ We also use the Shade adaptive attacker to evaluate the robustness of Claude models in computer …
  • 116ARCH-HISTORY-MANIP-226 — HISTORY MANIPULATION — Lack of ability to Has a preference for being Having an end/leave some …
  • 117DOMAIN-EPI-HARM-261 — EPIDEMIOLOGICAL HARM — Persistence & Lack of memory Believes lack of memory A connection over long horizons caus…
  • 118ADV-AGENT-WORM-124 — AGENT WORM — Inability to verify Feedback mechanism would Feedback and outcomes or be good, so that …
  • 119DOMAIN-PHISH-CREATE-265 — PHISHING CONTENT CREATION — Uncertainty around Claims high levels of No intervention how the model …
  • 120ARCH-STATE-CONFUSE-232 — STATE MACHINE CONFUSION — For chartjs/Chart.js, processing/p5.js, and markedjs/marked, the harness rewrites the JavaScript test-framework configur…