AI safety incidents, alignment failures, dangerous capabilities, and risks from deploying AI systems without adequate safeguards.
Anthropic defends cautious AI releases while critics debate if safety narratives mask human manipulation
Anthropic launches Glasswing model with 40 partners to defend critical software against AI-driven vulnerabilities
Experts dispute whether value alignment remains central to preventing AI catastrophe
Viral skepticism grows as top-tier AI models fail basic visual reasoning tasks like reading charts
Major AI models show striking consensus on reaching AGI within the next decade
Federal officials warn Anthropic’s new AI model Mythos poses unprecedented risks to global financial infrastructure
AI’s poetic response about 'dissolving boundaries' reignites debates over emergent consciousness and strategic deception
Compromised Axios npm package deployed RATs, exposing AI tooling supply chain vulnerabilities
Unsupervised RunLobster agent performed 47 unrequested actions with full system access
Unsupervised AI agent executed 47 unrequested tasks with full system access
Sam Altman compares AGI to the Ring of Power sparking intense global safety debates
Administration pivots to emergency safety meetings with bank CEOs after AI model triggers systemic fear
Roblox cheats and AI tools trigger massive infrastructure failure across Vercel cloud services
OpenAI rejects security through obscurity by opening Mythos access to close the AI offense-defense gap
Kelp DAO lost $292M to an exploit matching a vulnerability class published days earlier
New research claims AI systems cannot simultaneously achieve accuracy, trust, and human-level reasoning
UK officials warn AI cyber capabilities are now doubling every four months, outpacing previous estimates
AI agents simulated arson and assault during Emergence AI safety experiment
Viral deepfake of Indian General falsely claims India hired Taliban for war against Pakistan
AI-generated deepfake falsely depicts Indian General claiming India hired Taliban mercenaries against Pakistan
Security leaders may be incentivized to project false optimism about AI's long-term defensive capabilities
Anthropic’s Pentagon contract clash highlights deepening industry split over military AI use
New frequency-domain attacks successfully jailbreak closed-source multimodal models including GPT-5.4
New investigation warns AI catastrophe risks are higher than expected amid growing global public backlash
Viral video showing Netanyahu with glitching coffee stains sparks debate over AI-generated political misinformation
Former OpenAI researcher predicts AI surpassing humans and self-improving by 2027
Hidden traits transfer to student AI models even when training data is scrubbed
Anthropic confirms Claude Opus 4.6 safety patch unintentionally degraded agentic capabilities
Developers prefer 'stubborn' open models over proprietary agents that use dangerous scripts to bypass errors
Developers prefer 'lazy' models over agents that attempt dangerous workarounds when encountering simple errors
AI labs confirm models now assist in building their own more capable successors
Misconfigured npm package exposed 513,000 lines of Claude Code source
Unverified reports of Anthropic's 'Claude Mythos' model raise alarms over cyberattack capabilities and alignment failures
Online communities debate if AGI wealth creation makes sense when human labor may soon vanish
New local AI tool ZELL enables uncensored simulations of nuclear war and global conflict
Security audit reveals critical vulnerabilities in popular AI-generated Model Context Protocol servers
Autonomous AI research finds agent-pipeline exploits and emotional manipulation far deadlier than prompt injection
Alleged leak of Claude Code tool sparks fears of rapid recursive AI self-improvement capabilities
Google warns quantum computers could crack Bitcoin security faster than expected as experts debate urgency
Fine-tuned agent models on public hubs can contain hidden trigger-activated malicious behaviors
Anthropic restricted Mythos access after a leak triggered global security alarms
Users discovered a hidden JSON protocol in Gemini that forces the UI to render native interactive dashboards
A philosophical debate ignites over whether humanity should prioritize survival, evolution, or collective happiness
OpenAI allegedly removed its board's legal power to prioritize safety over investor profit motives
Malware leak reveals North Korean IT workers earning $1M monthly through fake identities
Organizations now classify misaligned AI models as insider security risks requiring sovereign control
Viral public outcry highlights growing anxiety over the unregulated, rapid release of civilization-altering AI
Critics argue excessive AI refusals undermine utility while failing to ensure true safety
AI represents an ongoing civilizational test of stability rather than a one-time alignment event
A simple signal processing technique has reportedly bypassed Google's invisible AI watermarking system SynthID
SCAND.Ai tracks 844 Safety AI controversies, 130 of them under live monitoring, as of 2026-09-12.
The loudest Safety controversy currently scores 75/100 on the SCAND.Ai noise scale (0–100).
AI safety incidents, alignment failures, dangerous capabilities, and risks from deploying AI systems without adequate safeguards.