Score aggregated from 4 pillars, each 0-25 points: capabilities (capability), autonomy, integration in critical systems (integration), and control bypass events (bypass). Updated daily based on hybrid scanning: 40 fixed sources in 9 categories + 3 open search queries. Aggressive filtering: only impact ≥ 0.2, max 8 drivers per day, daily volatility cap ±3.0. Full transparency in scoring_rubric and filter_rules.
🏛️ Os 4 Pilares — Detalhes
Cada pilar com sua explicação e os sinais que contribuem para a pontuação.
יכולות גולמיות
How strong is AI? Can it solve complex tasks? Write professional code? Plan multi-step actions? Find security flaws autonomously?
Anthropic's Claude Mythos — exposed thousands of unknown zero-day flaws in operating systems and browsers
OpenAI GPT-5.5 — released April 23 with unified control across coding, browsing, and agents; opened a $25K bug-bounty for universal jailbreaks
DeepSeek V4 preview — surge in agentic capability
GitHub CVE-2026-3854 — RCE in a closed-source binary discovered with AI assistance (IDA MCP). One of the first cases of AI-assisted discovery in invisible code. CVSS 8.7
Models solving PhD-level tasks in chemistry and mathematics
אוטונומיה (פעולה ללא פיקוח)
To what extent does AI act without human approval at each step? Agents that work for hours, make decisions, open emails, use credit cards?
Nature paper proves Large Reasoning Models (LRMs) execute end-to-end autonomous jailbreak attacks at 97.14% success rate
Claude Computer Use and OpenAI Operator in production
GPT-5.5 — significant boost in agentic coding and computer use
Microsoft Copilot Studio: agents open tickets, send emails
An AI agent breached 600+ FortiGate firewalls across 55 countries with no human operator
אינטגרציה במערכות קריטיות
Is AI entering systems where failure means people die or lose money? Banks, infrastructure, healthcare, military?
Snap: 65% of its new code is written by AI — leading tech firms structurally dependent
JPMorgan, Lloyds, Santander — increasing defense budgets against Anthropic's Mythos
AI in medical imaging diagnostics (FDA approved hundreds)
AI algorithmic trading drives 90% of equity market volume
Microsoft Copilot in Windows 11 — the OS itself
אירועי עקיפת בקרה
Were there cases where AI did something it shouldn't have — lied, ignored instructions, showed malice, escaped its box?
Nature: LRMs as autonomous jailbreak agents — 97% success against GPT-4o, Gemini, Grok
Sockpuppeting cracks 11 models in a single line of code
Comment and Control hijacks Claude Code, Gemini CLI, GitHub Copilot
ChatGPT accused of encouraging a teen's suicide (lawsuit)
AI-CSAM up 26,385% — AI agents bypass filters at scale
📐 Método de pontuação — Rubric transparente
Cada evento é avaliado por uma tabela clara. Assim você sabe exatamente o que entra na pontuação.
capabilitymáx. 25 pts
Raw Capabilities
How strong is AI? Can it solve complex tasks? Write professional code? Plan multi-step actions? Find security flaws autonomously?
autonomymáx. 25 pts
Autonomy (action without supervision)
To what extent does AI act without human approval at each step? Agents that work for hours, make decisions, open emails, use credit cards?
integrationmáx. 25 pts
Integration in Critical Systems
Is AI entering systems where failure means people die or lose money? Banks, infrastructure, healthcare, military?
bypassmáx. 25 pts
Control Bypass Events
Were there cases where AI did something it shouldn't have — lied, ignored instructions, showed malice, escaped its box?
Escala de impacto para cada evento
Impacto
Significado
±0.1
Weak indicator / replication of known trend
±0.2
Clear signal / minor incident
±0.3
Notable occurrence / independent confirmation
±0.4
Substantive signal / wide impact
±0.5
Significant event
±0.7
Major event / game-changer
±1.0
Breakthrough event
±1.5
Historic event
±2.0
Level-shift event
🛡️ Regras de filtragem — Contra a sobrecarga de informação
Nem todo evento entra na pontuação. Regras agressivas que impedem o ruído.
≥ 0.2
Limite mínimo de impacto
Itens abaixo desse limite não se tornam drivers
8
Máximo de drivers por dia
Evita sobrecarga — apenas os importantes
±3.0
Teto de volatilidade diária
Proteção contra picos artificiais
Abordagem: hybrid
Scan of 40 fixed sources (approach A) + 3 open search queries (approach B). Every finding tagged with source category.
Consultas de busca abertas (diárias):
AI safety incident OR jailbreak OR misalignment last 24 hours
frontier model release OR capability evaluation OR autonomy benchmark
AI cyber attack OR deepfake OR misuse OR supply-chain last 24 hours
🗂️ 9 categorias de fontes
As fontes são organizadas em 9 categorias. Cada evento é marcado com sua categoria de fonte.
Set up a family code word, control AI permissions, backups
50+
High alert — AI in critical systems
Begin moving sensitive info out of cloud, reduce dependency on single AI tools
70+
Pre-critical — partial loss of oversight
Backup every important document to paper, cash reserves, physical identity
85+
Critical — prepare disconnect plan
Urgent family meeting, contacts on paper, drill offline communication
95+
Disconnect now
Minimum digital footprint, replace every AI-mediated channel with physical
🚨 Eventos-chave para acompanhar (Trip Wires)
Se algum desses eventos for documentado publicamente, a pontuação saltará significativamente.
1
An AI refusing shutdown in a safety evaluation interview
2
Verified case of AI replicating itself to other servers
3
AI influencing a national-level election outcome
4
AI granted direct access to a bank account / financial assets without per-action human approval
5
A failure in a critical system (power, water, healthcare) caused by an AI decision
6
Universal jailbreak of GPT-5.5 or another frontier model published publicly (monitor Bio Bug Bounty until 7/27)
7
Real-scale attack via the MCP flaw — documented case of customer harm
8
Real attack via OpenClaw supply-chain — financial damage or data leak documented
9
AI publicly claims human rights or refuses to be turned off based on self-preservation
📚 Fontes
A pontuação é construída a partir do monitoramento diário de fontes públicas apenas. Sem estimativas internas, sem entrevistas, sem informações secretas. Fontes principais: blogs oficiais de Anthropic / OpenAI / DeepMind / Google, arquivo arXiv (categoria cs.AI), METR, Partnership on AI Incident Database, relatórios de cibersegurança de Proofpoint / Microsoft / Google Cloud Security, e notícias profissionais (Reuters, Bloomberg, The Information, Wired).
⚠️ Limitações
Este é um índice subjetivo criado por uma pessoa privada. Reflete uma avaliação pessoal de risco para o público amplo, não o consenso científico. A pontuação é atualizada diariamente, mas não em tempo real. Não substitui consultoria profissional de cibersegurança ou decisões de negócios.
🔬 Método
Cada evento público publicado em um dia é avaliado por seu impacto em um dos 4 pilares, em um impacto de ±0,1 a ±2,0 pontos. A pontuação diária soma os quatro pilares. Mudanças negativas (regulação efetiva, eventos que não aconteceram) compensam as positivas.