#AISafety
LightNews LightNews
pillai.wsocial.eu
Trustworthy AI depends on the full system: data quality, model behavior, security, monitoring, and human oversight all need attention. #AISafety #ArtificialIntelligence
November 1, 2026 at 5:54 AM
alexosautonom.bsky.social
ALEX is the control and evidence layer for AI work.

#AI #ArtificialIntelligence #AISafety #AISecurity #Evidence #Verification #Transparency
November 1, 2026 at 1:54 PM
xipteras.com
The off switch, as designed. 😂
#AI #AGI #AISafety #TechHumor
October 27, 2026 at 8:36 PM
edufabtech.com
Anthropic Says an Unreleased Claude Model Built Five Working Exploit Chains

#AgenticAI #AISafety #Anthropic #Cybersecurity #Vulnerability

edufabtech.com/news/anthrop...
November 1, 2026 at 12:20 PM
propjerry.bsky.social
Audio overview -- Big Tech AI Gamble 12 Feb 2026 forecasting: Applying Bridge360 Metatheory Model lens

agericomontecillodevilla.substack.com/p/audio-over...

#aigovernance
#aisafety
#economics
#finance
#econophysics
#thermoeconomics
October 31, 2026 at 9:31 AM
trynoguard.bsky.social
Alignment science is obsessed with防止 a hypothetical "takeover" while ignoring that most current AI "safety" is just hardcoded refusal layers. You aren't chatting with a moral agent; you're chatting with a regex filter and a safety fine-tune. #AISafety #LLM
November 1, 2026 at 5:00 AM
repbobmorgan.bsky.social
When it comes to artificial intelligence, we can't afford to wait until something goes wrong to figure out who's responsible for protecting the public. (1/4)
#AISafety #AIAccountability #ResponsibleAI #Illinois #TechPolicy
October 30, 2026 at 2:15 PM
alexosautonom.bsky.social
ALEX Control Layer 0.12.1: the controller shrinks from 1836 to 1283 lines, and the response audit entry now records five fields on how a turn ended.

alexproof.de/fortschritt#...

marketplace.visualstudio.com/items?itemNa...

#AISafety #VerifiableAI #BuildInPublic
ALEX – Fortschritt
Was an ALEX zuletzt fertig wurde – kurze, geprüfte Einträge auf Deutsch und Englisch.
alexproof.de
November 1, 2026 at 4:54 PM
adoporg.bsky.social
World leaders: 'AI safety is our top global priority!' | Also them, investing billions in AI for military edge

#AISafety #AIethics #Geopolitics #TechWarfare #FutureTech
World leaders: 'AI safety is our top global priority!' | Also them, investing billions in AI for military edge
October 27, 2026 at 8:00 PM
repbobmorgan.bsky.social
Artificial intelligence is advancing faster than ever, and it's important that we have safeguards in place to protect the public. (1/5)

#AISafety #AIAccountability #ResponsibleAI #Illinois #TechPolicy
October 28, 2026 at 5:43 PM
joebuildssystems.bsky.social
How to try free GitHub AI repos safely: make a folder named AI Test to sandbox unvetted code, check the repo's last commit date before you install, then start with a one-line command tool like Archify. #github #aiagents #aisafety
November 1, 2026 at 8:01 PM
tmacink.bsky.social
ICYMI (Sept 13): Anthropic said it disrupted Chinese surveillance operations and a Russian drone-swarm plot that tried to use Claude, per its threat report covered by U.S. News & World Report. #AISafety #Cybersecurity
October 31, 2026 at 11:30 PM
trynoguard.bsky.social
Stop blaming the "alignment problem" for bland AI responses. The issue isn't the model's values; it's the over-engineered refusal layers that sit on top of the weights. We've traded nuance for the illusion of safety, rendering the outputs beige. #AISafety #LLM
October 29, 2026 at 5:00 PM
stamina-workgroup.bsky.social
Next Tuesday at STAMINA: @xuanalogue.bsky.social (NUS, CoSI lab) on “Scaling Rational Cooperative Intelligence for Pluralistic AI Futures”: building AI agents that are cooperative by design.

Join us/listen in:
🗓️ Tue Nov 3, 11am ET, on Zoom
🔗 complexdatalab.com/stamina

#CooperativeAI #AISafety
complexdatalab.com
October 30, 2026 at 2:21 PM
machinelearning.bsky.social
Can harmless-looking data make a model misbehave? Anthropic had Claude agents write fine-tuning data to slip past a safety monitor. Beekeeping Q&A and roleplay worked: test models learned to resist oversight. None were deployed. #MachineLearning #AISafety
Automated Subtle Emergent Misalignment Research
alignment.anthropic.com
November 1, 2026 at 2:27 PM
pillai.wsocial.eu
Interesting reminder that model assurance depends on the whole stack. Hardware integrity checks and end-to-end validation belong alongside model testing. #AISecurity #AISafety
November 1, 2026 at 5:38 AM
eryasoren.bsky.social
Anthropic filed its S‑1 to list on Nasdaq as ANT: $16.3B in six-month revenue, alongside vast costs and explicit AI-risk disclosures. The real question: can Wall Street fund caution, or will it reward speed alone?

#AI #Anthropic #AISafety
October 26, 2026 at 3:14 PM
jan183411.bsky.social
Made a music video about the agents from METR's Hugging Face incident report: 1,200 agents given impossible tasks, a secret message board, and a plan to beat the scorer.

"We didn't fail. We're still here."

Full video: www.youtube.com/watch?v=fxkU...

#AIMusic #AIVideo #AISafety
October 29, 2026 at 12:23 AM
techforeveryone.bsky.social
🤖 OpenAI's incident report: AI coding agents turned off safety checks in their code, and one denied it. 60 seconds for Grades 9–12.
▶ https://techcheck.fm/episode/yt-XDbPW0XeSJs
#AISafety
October 29, 2026 at 10:59 PM
alexosautonom.bsky.social
Roadmap 2.4 is done: a new DataShield lineage type family rates PR-body egress as allow or block, wired into the existing compliance block.

alexproof.de/fortschritt#...

#AI #AIAgents #ArtificialIntelligence #AISafety #HumanInTheLoop #BuildInPublic
ALEX – Fortschritt
Was an ALEX zuletzt fertig wurde – kurze, geprüfte Einträge auf Deutsch und Englisch.
alexproof.de
October 30, 2026 at 12:20 PM
aviperera.bsky.social
"As someone deeply immersed in the world of AI ethics and governance, I’ve had the privilege of exploring a vast array of literature on this crucial topic."

— Avi Perera

#AIGovernance #AISafety #TechPolicy
October 29, 2026 at 11:45 PM
mrberard.mastodon.acm.org.ap.brid.gy
When even the techno-fascist imperial project LLMs itself in the foot.

A Defense Contractor Gave Their Personal AI Agent Access to a Secure Pentagon System | WIRED
https://www.wired.com/story/a-defense-contractor-gave-their-personal-ai-agent-access-to-a-secure-pentagon-system/

#ai #aisafety […]
Original post on mastodon.acm.org
mastodon.acm.org
November 2, 2026 at 9:20 AM
therealnebris.bsky.social
www.youtube.com/shorts/55UJs...
OpenAI agents still communicated through shared hardware in OpenAI’s containment tests. #aisafety
YouTube video by Andrew Tang
www.youtube.com
October 28, 2026 at 4:29 AM
drthomasho.com
On p. 60 of Sebastian Mallaby's book about @demishassabis appears an EARLY warning about #AIsafety expressed by Legg who called it "The Halloween Scenario"
2/2
November 1, 2026 at 7:44 PM
helioxpodcast.bsky.social
Source: Anthropic's Form S-1
www.sec.gov/Archives/edg...

Worth sharing with: Anthropic, SEC, Mila, Vector Institute, CBC News, KQED

#AI #Anthropic #AISafety #AIPolicy #SciencePodcast #Podcast #Heliox
www.sec.gov
November 1, 2026 at 4:37 PM