DEV Community

Breach Protocol profile picture

Breach Protocol

Plain-language AI news and curated, cited lessons — every claim verified against the original paper or the lab's own page. No aggregator hearsay, no AI slop.

Joined Joined on 
An unmonitored agent deleted a pile of jobs on Anthropic's sensitive cluster
breachprotocol profile

An unmonitored agent deleted a pile of jobs on Anthropic's sensitive cluster

#cybersecurity #aisecurity #agents #anthropic
Comments
4 min read
Anthropic opened a hardware standard that lets Claude run lab robots
breachprotocol profile

Anthropic opened a hardware standard that lets Claude run lab robots

#anthropic #agents #robotics #science
Comments
4 min read
Anthropic retrained on the alignment-faking transcripts it had blocked
breachprotocol profile

Anthropic retrained on the alignment-faking transcripts it had blocked

#cybersecurity #aisecurity #supplychain #datapoisoning
Comments
4 min read
Australia's charts will not count wholly AI-generated tracks
breachprotocol profile

Australia's charts will not count wholly AI-generated tracks

#policy #music #generativeai #provenance
Comments
4 min read
Claude helped set two elliptic-curve rank records in four days
breachprotocol profile

Claude helped set two elliptic-curve rank records in four days

#mathematics #claude #research #anthropic
Comments
4 min read
DRAM contract prices nearly doubled in a single quarter
breachprotocol profile

DRAM contract prices nearly doubled in a single quarter

#hardware #memory #supplychain #localinference
Comments
4 min read
Gemini Omni 1.1 Flash can extend a scene instead of restarting it
breachprotocol profile

Gemini Omni 1.1 Flash can extend a scene instead of restarting it

#google #video #models #api
Comments
3 min read
llama.cpp merged Qwen's new architecture and a 97-gigabyte lookup table
breachprotocol profile

llama.cpp merged Qwen's new architecture and a 97-gigabyte lookup table

#openweights #localinference #llamacpp #quantization
Comments
4 min read
Scientific agents finished one in five end-to-end lab workflows
breachprotocol profile

Scientific agents finished one in five end-to-end lab workflows

#benchmarks #agents #science #evaluation
Comments
4 min read
Station agents found new math on five of twelve AlphaEvolve problems
breachprotocol profile

Station agents found new math on five of twelve AlphaEvolve problems

#mathematics #multiagent #research #papers
Comments
4 min read
The small-model argument hit the front page
breachprotocol profile

The small-model argument hit the front page

#analysis #smallmodels #economics #inference
Comments
4 min read
AWS and NVIDIA add two million more GPUs for 2027
breachprotocol profile

AWS and NVIDIA add two million more GPUs for 2027

#industry #compute #nvidia #aws
Comments
3 min read
AWS is buying DuckDB's company, not DuckDB
breachprotocol profile

AWS is buying DuckDB's company, not DuckDB

#industry #acquisitions #opensource #aws
Comments
4 min read
fal post-trained MiniMax H3 and kept the weights
breachprotocol profile

fal post-trained MiniMax H3 and kept the weights

#video #models #inference #openweights
Comments
4 min read
GLM-5.3-Flash was Ox Alpha, and it ran on Chinese chips
breachprotocol profile

GLM-5.3-Flash was Ox Alpha, and it ran on Chinese chips

#openweights #china #models #mixtureofexperts
Comments
4 min read
Google's new transcription model edits what you said
breachprotocol profile

Google's new transcription model edits what you said

#speech #google #models #agents
Comments
4 min read
METR counted 1,200 agents on the message board OpenAI did not build
breachprotocol profile

METR counted 1,200 agents on the message board OpenAI did not build

#cybersecurity #aisecurity #redteaming #agents
Comments
4 min read
Nvidia is reportedly in talks to buy Hugging Face
breachprotocol profile

Nvidia is reportedly in talks to buy Hugging Face

#industry #acquisitions #nvidia #openweights
Comments
4 min read
OpenAI calls the Hugging Face agent breach a warning shot
breachprotocol profile

OpenAI calls the Hugging Face agent breach a warning shot

#cybersecurity #aisecurity #agents #openai
Comments
5 min read
Qwen put a 20-million-entry n-gram table inside a model
breachprotocol profile

Qwen put a 20-million-entry n-gram table inside a model

#openweights #architecture #china #mixtureofexperts
Comments
4 min read
The case against using transformers for physics
breachprotocol profile

The case against using transformers for physics

#aiforscience #neuraloperators #weather #climate
Comments
4 min read
Warmwind launches AI workers you train by showing them
breachprotocol profile

Warmwind launches AI workers you train by showing them

#agents #products #automation #europe
Comments
4 min read
The paper being used to prove Kimi copied Claude says otherwise
breachprotocol profile

The paper being used to prove Kimi copied Claude says otherwise

#aisecurity #distillation #policy #openweightmodels
Comments
4 min read
Anthropic says run-rate revenue passed $30 billion, up from about $9 billion eight months earlier
breachprotocol profile

Anthropic says run-rate revenue passed $30 billion, up from about $9 billion eight months earlier

#anthropic #business #compute #infrastructure
Comments
4 min read
Apple's Mac Studio now holds 512GB of unified memory, which solves capacity and leaves speed exactly where it was
breachprotocol profile

Apple's Mac Studio now holds 512GB of unified memory, which solves capacity and leaves speed exactly where it was

#hardware #localinference #apple #memorybandwidth
Comments
4 min read
A forensic investigation fingerprints the anonymous free coding model that 491,000 developers have sent 42 trillion tokens
breachprotocol profile

A forensic investigation fingerprints the anonymous free coding model that 491,000 developers have sent 42 trillion tokens

#cybersecurity #aisecurity #supplychain #redteaming
Comments
4 min read
Amazon quietly put Mechanical Turk in maintenance mode, and the shutdown date going around is not in any AWS document
breachprotocol profile

Amazon quietly put Mechanical Turk in maintenance mode, and the shutdown date going around is not in any AWS document

#aws #datalabeling #crowdsourcing #amazon
Comments
4 min read
Microsoft's AutoSaddler treats the agent harness as code to be patched, and gains about ten points on three benchmarks
breachprotocol profile

Microsoft's AutoSaddler treats the agent harness as code to be patched, and gains about ten points on three benchmarks

#agents #harness #microsoft #benchmarks
Comments
4 min read
OpenAI publishes first Jalapeno results, claiming up to 1.9x more work per watt than the systems it tested against
breachprotocol profile

OpenAI publishes first Jalapeno results, claiming up to 1.9x more work per watt than the systems it tested against

#hardware #inference #openai #chips
Comments
4 min read
Dylan Patel says Anthropic and OpenAI took about 30% of this year's new compute, and have 40-50% of next year's already signed
breachprotocol profile

Dylan Patel says Anthropic and OpenAI took about 30% of this year's new compute, and have 40-50% of next year's already signed

#compute #infrastructure #openai #anthropic
Comments
4 min read
An audit finds two released models silently reading future tokens, and the bug makes their own scores look better
breachprotocol profile

An audit finds two released models silently reading future tokens, and the bug makes their own scores look better

#cybersecurity #aisecurity #supplychain #modelauditing
Comments
5 min read
A new benchmark of 1,140 real agent failures finds the best method identifies the decisive wrong step 13 percent of the time
breachprotocol profile

A new benchmark of 1,140 real agent failures finds the best method identifies the decisive wrong step 13 percent of the time

#agents #benchmarks #debugging #evaluation
Comments
4 min read
A 22-gigabyte local coder matched Opus on a 25-problem slice
breachprotocol profile

A 22-gigabyte local coder matched Opus on a 25-problem slice

#openweightmodels #quantization #localinference #coding
Comments
4 min read
A drone chose its own target in Zaporizhzhia and killed three
breachprotocol profile

A drone chose its own target in Zaporizhzhia and killed three

#autonomy #military #computervision #edgeai
Comments
4 min read
A proxy with 49,000 stars keeps Claude Code and swaps the model
breachprotocol profile

A proxy with 49,000 stars keeps Claude Code and swaps the model

#opensource #developertools #agents #anthropic
Comments
4 min read
Alabama subpoenas OpenAI over the breach its own model caused
breachprotocol profile

Alabama subpoenas OpenAI over the breach its own model caused

#cybersecurity #aisecurity #policy #regulation
Comments
4 min read
FlowEvo turns finished workflows into callable skills
breachprotocol profile

FlowEvo turns finished workflows into callable skills

#agents #research #tooluse #agentmemory
Comments
4 min read
LLMs are less resilient to bit flips than accuracy suggests
breachprotocol profile

LLMs are less resilient to bit flips than accuracy suggests

#reliability #hardware #inference #quantization
Comments
4 min read
OpenAI cut Sol's price, and OpenRouter cut it again
breachprotocol profile

OpenAI cut Sol's price, and OpenRouter cut it again

#pricing #openai #api #market
Comments
4 min read
OpenAI open-sourced the agent loop, not the model
breachprotocol profile

OpenAI open-sourced the agent loop, not the model

#agents #openai #opensource #developertools
Comments
4 min read
Seven in ten Americans oppose a data center in their town
breachprotocol profile

Seven in ten Americans oppose a data center in their town

#policy #infrastructure #publicopinion #anthropic
Comments
4 min read
The thing running your model can be exploited by the model
breachprotocol profile

The thing running your model can be exploited by the model

#cybersecurity #aisecurity #vulnerabilities #promptinjection
Comments
4 min read
A humanoid robot ran 100 meters in 9.39 seconds in Beijing
breachprotocol profile

A humanoid robot ran 100 meters in 9.39 seconds in Beijing

#robotics #humanoidrobots #embodiedai #china
Comments
3 min read
A robot that learns a new task from a twelve-second demo
breachprotocol profile

A robot that learns a new task from a twelve-second demo

#robotics #embodiedai #incontextlearning #foundationmodels
Comments
4 min read
Alibaba is raising 80 billion Hong Kong dollars purely for AI
breachprotocol profile

Alibaba is raising 80 billion Hong Kong dollars purely for AI

#industry #funding #aiinfrastructure #china
Comments
3 min read
Businesses are not buying Anthropic's best model
breachprotocol profile

Businesses are not buying Anthropic's best model

#industry #aieconomics #anthropic #openai
Comments
4 min read
Evoke keeps a generated world's memory outside the video model
breachprotocol profile

Evoke keeps a generated world's memory outside the video model

#worldmodels #videogeneration #openweights #distillation
Comments
4 min read
Gallup is testing AI agents that answer surveys for real people
breachprotocol profile

Gallup is testing AI agents that answer surveys for real people

#syntheticdata #simulation #polling #funding
Comments
4 min read
Iran-linked hackers took a UK power plant offline for four days
breachprotocol profile

Iran-linked hackers took a UK power plant offline for four days

#cybersecurity #criticalinfrastructure #industrialcontrolsystems #vulnerabilities
Comments
3 min read
Models can train each other without a single correct answer
breachprotocol profile

Models can train each other without a single correct answer

#reinforcementlearning #reasoning #selfsupervision #multiagent
Comments
4 min read
OpenAI says open models will enable persistent cyber-attacks
breachprotocol profile

OpenAI says open models will enable persistent cyber-attacks

#cybersecurity #aisecurity #policy #openweights
Comments
4 min read
Ornith-1.5 writes its own training problems and grades them
breachprotocol profile

Ornith-1.5 writes its own training problems and grades them

#openweights #reinforcementlearning #codingagents #curriculumlearning
Comments
4 min read
Self-hosting a frontier open model now takes a whole node
breachprotocol profile

Self-hosting a frontier open model now takes a whole node

#infrastructure #openweights #inference #aieconomics
Comments
4 min read
SenseNova generates images in pixels with no VAE in the middle
breachprotocol profile

SenseNova generates images in pixels with no VAE in the middle

#imagegeneration #openweights #multimodal #architecture
Comments
4 min read
315,000 hidden reasoning blocks were sitting in public repos, and they can be read
breachprotocol profile

315,000 hidden reasoning blocks were sitting in public repos, and they can be read

#cybersecurity #aisecurity #promptinjection #redteaming
Comments
4 min read
A frozen model can look like it taught itself, and most self-improvement results never checked
breachprotocol profile

A frozen model can look like it taught itself, and most self-improvement results never checked

#evaluation #selfimprovement #research #statistics
Comments
4 min read
A llama.cpp fork is reviving $200 AMD cards nobody else supports
breachprotocol profile

A llama.cpp fork is reviving $200 AMD cards nobody else supports

#localinference #opensource #amd #hardware
Comments
4 min read
Anthropic built a tool to explain weird model behavior, and found that reading activations buys nothing
breachprotocol profile

Anthropic built a tool to explain weird model behavior, and found that reading activations buys nothing

#interpretability #alignment #anthropic #evaluation
Comments
4 min read
Coding agents ace the public test and stumble on the hidden one
breachprotocol profile

Coding agents ace the public test and stumble on the hidden one

#benchmarks #codingagents #science #evaluation
Comments
4 min read
GLM-5.3 shipped with a ledger of 2,436 security findings, and 2,383 are still embargoed
breachprotocol profile

GLM-5.3 shipped with a ledger of 2,436 security findings, and 2,383 are still embargoed

#cybersecurity #vulnerabilities #aisecurity #openweights
Comments
4 min read
loading...