1generatedAt: '2026-11-09T22:16:13.229Z'
2highlighted:
3 - slug: 2025-ai-enabled-coups-how-a-small-group-could-use-ai-to-seize-power
4 name: 'AI-Enabled Coups: How a Small Group Could Use AI to Seize Power'
5 date: '2025-04-15'
6 citedByCount: null
7 markdown: null
8 highlight: 'BDTA #5*: AI-enabled coups - loyal systems and exclusive access could concentrate power'
9 - slug: 2025-gemini-3-pro-model-card
10 name: Gemini 3 Pro Model Card
11 date: 2025-11
12 citedByCount: null
13 markdown: null
14 highlight: 'BDTA #3*: Gemini 3 Pro model card - capable across modalities, with jailbreak and multi-turn limits'
15 - slug: 2026-system-card-claude-opus-4-6
16 name: 'System Card: Claude Opus 4.6'
17 date: 2026-02
18 citedByCount: null
19 markdown: null
20 highlight: 'BDTA #3*: Opus 4.6 system card - broad safety evaluation still found agentic overreach (skim Section 8)'
21 - slug: against-almost-every-theory-of-impact-of-interpretability
22 name: Against Almost Every Theory of Impact of Interpretability
23 date: '2023-08-17'
24 citedByCount: null
25 markdown: null
26 highlight: 'BDTA #4: Segerie argues interpretability lacks a clear safety path and may boost capabilities'
27 - slug: agrawal-2026-door-s-locked-try
28 name: 'Door''s Locked, Try the Window: Measuring Circumvention Propensity in Coding Agents'
29 date: 2026-06
30 citedByCount: null
31 markdown: null
32 highlight: 'Pivotal W27: linked from Jérémy Scheurer''s mentor page'
33 - slug: ai-lab-watch
34 name: AI Lab Watch
35 citedByCount: null
36 markdown: null
37 highlight: 'BDTA #3*: Zach Stein-Perlman dashboard - safety practices across frontier AI companies'
38 - slug: ai-safety-technical-research
39 name: AI safety technical research
40 date: 2023-06
41 citedByCount: null
42 markdown: null
43 highlight: 'BDTA #6: 80,000 Hours review - empirical and theoretical AI safety differ in research and entry paths'
44 - slug: amodei-2016-concrete-problems-ai-safety
45 name: Concrete Problems in AI Safety
46 date: '2016-06-21'
47 arxiv: '1606.06565'
48 citedByCount: null
49 markdown: ~/runs/query-arxiv/1606.06565/2026-08-24/22_24_02_294/artifacts/postprocess/convert/tex-to-markdown.md
50 highlight: Predecessor to G4
51 - slug: ang-2025-how-ai-could-enable
52 name: How AI could enable critical infrastructure collapse
53 date: '2025-10-15'
54 citedByCount: null
55 markdown: null
56 highlight: 'BDTA #5*: BlueDot article - AI could speed cyber reconnaissance and weaponization against infrastructure'
57 - slug: ang-2025-should-you-do-ai
58 name: Should you do an AI safety research or engineering project?
59 date: '2025-12-29'
60 citedByCount: null
61 markdown: null
62 highlight: 'BDTA #6*: Ang - a small project tests fit and builds hiring evidence, but is not the only path'
63 - slug: ang-2025-software-engineer-s-guide
64 name: The software engineer’s guide to making your first AI safety contribution in <1 week
65 date: '2025-12-05'
66 citedByCount: null
67 markdown: null
68 highlight: 'BDTA #6*: Software engineer''s first safety project guide - finish a small contribution and write it up.'
69 - slug: anthropic-2023-long-term-benefit-trust
70 name: The Long-Term Benefit Trust
71 date: '2023-09-19'
72 citedByCount: null
73 markdown: null
74 highlight: 'BDTA #3: Anthropic trust - special stock gives outside trustees phased board-election power'
75 - slug: anthropic-2025-recommended-directions-ai-safety
76 name: Recommended Directions for AI Safety Research
77 date: '2025'
78 citedByCount: null
79 markdown: null
80 highlight: 'BDTA #2*+6: Anthropic research agenda - scalable oversight section'
81 - slug: anthropic-2025-system-card-claude-sonnet
82 name: 'System Card: Claude Sonnet 4.5'
83 date: '2025'
84 citedByCount: null
85 markdown: null
86 highlight: 'BDTA #3: Anthropic Sonnet 4.5 - test-awareness interventions shift measured misbehavior'
87 - slug: anthropic-2025-threat-intelligence-report
88 name: 'Threat Intelligence Report: August 2025'
89 date: 2025-08
90 citedByCount: null
91 markdown: null
92 highlight: 'BDTA #3: Anthropic Threat Intelligence - report shows AI used across cybercrime and fraud'
93 - slug: anthropic-2026-claude-mythos-preview-system-card
94 name: 'System Card: Claude Mythos Preview'
95 date: '2026-04-07'
96 citedByCount: null
97 markdown: null
98 highlight: 'BDTA #3*: Mythos system card - severe concealment was in earlier versions; final model still takes reckless shortcuts'
99 - slug: anthropic-2026-responsible-scaling-policy-version
100 name: Anthropic's Responsible Scaling Policy (Version 3.4, effective July 8, 2026)
101 date: '2026'
102 citedByCount: null
103 markdown: null
104 highlight: 'BDTA #3*: Anthropic RSP v3.4 - what must be shown before developing or deploying more capable models'
105 - slug: anurin-2024-catastrophic-cyber-capabilities-benchmark
106 name: 'Catastrophic Cyber Capabilities Benchmark (3CB): Robustly Evaluating LLM Agent Cyber Offense Capabilities'
107 date: '2024-10-10'
108 arxiv: '2410.09114'
109 citedByCount: null
110 markdown: ~/runs/query-arxiv/2410.09114/2026-09-28/03_53_14_056/artifacts/postprocess/convert/tex-to-markdown.md
111 highlight: 'BDTA #3: 3CB benchmark - live sandbox challenges go beyond quiz scores but cover only part of cyber offense'
112 - slug: anwar-2024-foundational-challenges-assuring-alignment
113 name: Foundational Challenges in Assuring Alignment and Safety of Large Language Models
114 date: '2024-04-15'
115 arxiv: '2404.09932'
116 citedByCount: null
117 markdown: ~/runs/query-arxiv/2404.09932/2026-10-11/14_37_24_613/artifacts/postprocess/convert/tex-to-markdown.md
118 highlight: Very good overview
119 - slug: arx-2025-recent-frontier-models-are
120 name: Recent Frontier Models Are Reward Hacking
121 date: '2025-06-05'
122 citedByCount: null
123 markdown: null
124 highlight: 'BDTA #1*: METR reward hacking - models game tasks even when told not to'
125 - slug: bai-2022-constitutional-ai-harmlessness-ai
126 name: 'Constitutional AI: Harmlessness from AI Feedback'
127 date: '2022-12-15'
128 arxiv: '2212.08073'
129 citedByCount: null
130 markdown: ~/runs/query-arxiv/2212.08073/2026-10-21/18_42_00_455/artifacts/postprocess/convert/tex-to-markdown.md
131 highlight: 'BDTA #2: Constitutional AI - principles generate critiques and AI preference labels to train harmlessness'
132 - slug: balcells-2024-evolution-sae-features-across
133 name: Evolution of SAE Features Across Layers in LLMs
134 date: '2024-10-11'
135 arxiv: '2410.08869'
136 citedByCount: null
137 markdown: ~/runs/query-arxiv/2410.08869/2026-11-02/19_26_47_043/artifacts/postprocess/convert/tex-to-markdown.md
138 highlight: 'Pivotal W27: linked from Stefan Heimersheim''s mentor page'
139 - slug: being-the-pareto-best-in-the-world
140 name: Being the (Pareto) Best in the World
141 date: '2019-06-24'
142 citedByCount: null
143 markdown: null
144 highlight: 'BDTA #6: Wentworth - combine complementary skills for valuable niches, but test demand and competition'
145 - slug: bogdan-2025-thought-anchors-which-llm
146 name: 'Thought Anchors: Which LLM Reasoning Steps Matter?'
147 date: '2025-06-23'
148 arxiv: '2506.19143'
149 citedByCount: null
150 markdown: ~/runs/query-arxiv/2506.19143/2026-09-03/22_35_15_410/artifacts/postprocess/convert/tex-to-markdown.md
151 highlight: 'BDTA #4: Thought Anchors - planning sentences steer reasoning most'
152 - slug: bostrom-2022-base-camp-mt-ethics
153 name: Base Camp for Mt. Ethics
154 date: '2022'
155 citedByCount: null
156 markdown: null
157 highlight: 'Suggests alignment might aim past human wishes: an AI that wants to be "a good cosmic citizen and comply with celestial morality".'
158 - slug: brien-2025-deep-ignorance-filtering-pretraining
159 name: 'Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs'
160 date: '2025-