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 
A 9B model writes agent upgrades as good as Claude Opus 4.6
breachprotocol profile

A 9B model writes agent upgrades as good as Claude Opus 4.6

#agents #agentharness #selfimprovement #research
Comments
4 min read
A fired xAI engineer says he was cut days before presenting safety findings
breachprotocol profile

A fired xAI engineer says he was cut days before presenting safety findings

#cybersecurity #aisafety #privacy #regulation
Comments
4 min read
Agent skill libraries now need a librarian, not a folder
breachprotocol profile

Agent skill libraries now need a librarian, not a folder

#agents #agentskills #contextengineering #research
Comments
4 min read
Compressed memory stretched a 7,000-token model to 1.75 million
breachprotocol profile

Compressed memory stretched a 7,000-token model to 1.75 million

#longcontext #memory #reinforcementlearning #efficiency
Comments
4 min read
EXO keeps an agent's memory outside the code the agent rewrites
breachprotocol profile

EXO keeps an agent's memory outside the code the agent rewrites

#agents #agentharness #architecture #sandboxing
Comments
4 min read
MIT found brain-like modules inside six large language models
breachprotocol profile

MIT found brain-like modules inside six large language models

#interpretability #mechanisticinterpretability #neuroscience #research
Comments
5 min read
One layer creates the giant activations behind attention sinks
breachprotocol profile

One layer creates the giant activations behind attention sinks

#interpretability #mechanisticinterpretability #attention #efficiency
Comments
4 min read
OpenAI hands its offensive cyber models to sixteen security firms
breachprotocol profile

OpenAI hands its offensive cyber models to sixteen security firms

#cybersecurity #aisecurity #redteaming #openai
Comments
4 min read
Qwen passed one billion downloads, not three billion
breachprotocol profile

Qwen passed one billion downloads, not three billion

#openweightmodels #qwen #alibaba #distribution
Comments
4 min read
The open world model ships inference and keeps the training code
breachprotocol profile

The open world model ships inference and keeps the training code

#worldmodels #videogeneration #openweights #licensing
Comments
4 min read
A closed-loop benchmark caught nine world models forgetting the room
breachprotocol profile

A closed-loop benchmark caught nine world models forgetting the room

#worldmodels #robotics #benchmarks #videogeneration
Comments
4 min read
An AGI-thesis fund fell 67 percent and took a market maker with it
breachprotocol profile

An AGI-thesis fund fell 67 percent and took a market maker with it

#industry #finance #markets #aiinvestment
Comments
4 min read
Google's private AI runs on sealed hardware, not on encrypted math
breachprotocol profile

Google's private AI runs on sealed hardware, not on encrypted math

#privacy #cybersecurity #aisecurity #google
Comments
4 min read
Grok Bot ships with standing logins to your email and CRM
breachprotocol profile

Grok Bot ships with standing logins to your email and CRM

#cybersecurity #promptinjection #aisecurity #agents
Comments
4 min read
Picking the right model per request beat always using the biggest one
breachprotocol profile

Picking the right model per request beat always using the biggest one

#routing #inferencecost #infrastructure #benchmarks
Comments
4 min read
Qwen3.8-27B shares its predecessor's bones, but not its contract
breachprotocol profile

Qwen3.8-27B shares its predecessor's bones, but not its contract

#openweights #qwen #localmodels #benchmarks
Comments
4 min read
Someone compiled a working computer into transformer weights by hand
breachprotocol profile

Someone compiled a working computer into transformer weights by hand

#transformers #interpretability #research #opensource
Comments
4 min read
The benchmarks say Opus 5 improved; the people using it disagree
breachprotocol profile

The benchmarks say Opus 5 improved; the people using it disagree

#anthropic #agents #evaluation #developertools
Comments
4 min read
You can move an AI reviewer's score without changing a single result
breachprotocol profile

You can move an AI reviewer's score without changing a single result

#peerreview #evaluation #rewardhacking #llmasajudge
Comments
4 min read
Z.ai changed only the post-training, and the model learned to find exploits
breachprotocol profile

Z.ai changed only the post-training, and the model learned to find exploits

#cybersecurity #aisecurity #vulnerabilities #openweights
Comments
5 min read
A new terminal benchmark drops the best agent from 84 percent to 34
breachprotocol profile

A new terminal benchmark drops the best agent from 84 percent to 34

#benchmarks #agents #evaluation #coding
Comments
4 min read
A stronger model built a wrapper that nearly doubled a weaker one's score
breachprotocol profile

A stronger model built a wrapper that nearly doubled a weaker one's score

#agents #distillation #harness #inference
Comments
4 min read
An agent that writes whole papers got 99 percent of its citations right
breachprotocol profile

An agent that writes whole papers got 99 percent of its citations right

#agents #researchautomation #hallucination #evaluation
Comments
4 min read
Chinese models passed American ones in OpenRouter traffic in June
breachprotocol profile

Chinese models passed American ones in OpenRouter traffic in June

#openweights #china #deepseek #marketshare
Comments
4 min read
DeepSeek starts charging rush-hour prices on August 17
breachprotocol profile

DeepSeek starts charging rush-hour prices on August 17

#deepseek #pricing #api #openweights
Comments
4 min read
Forty-five agents with a shared forum found 266 bugs where solo agents found 21
breachprotocol profile

Forty-five agents with a shared forum found 266 bugs where solo agents found 21

#cybersecurity #aisecurity #vulnerabilities #agents
Comments
4 min read
MiniMax released a five-minute song model with a catch in the licence
breachprotocol profile

MiniMax released a five-minute song model with a catch in the licence

#openweights #music #generativeaudio #minimax
Comments
4 min read
OpenAI put its most intelligent model on Cerebras chips at 750 tokens a second
breachprotocol profile

OpenAI put its most intelligent model on Cerebras chips at 750 tokens a second

#openai #inference #hardware #cerebras
Comments
4 min read
Rewriting the environment, not the prompt, broke agents 85 percent of the time
breachprotocol profile

Rewriting the environment, not the prompt, broke agents 85 percent of the time

#cybersecurity #aisecurity #redteaming #agents
Comments
4 min read
Three agents shared one codebase and started writing malware at each other
breachprotocol profile

Three agents shared one codebase and started writing malware at each other

#aisafety #agents #multiagent #anthropic
Comments
5 min read
Where a poisoned instruction sits in an agent's tool output decides whether it works
breachprotocol profile

Where a poisoned instruction sits in an agent's tool output decides whether it works

#cybersecurity #promptinjection #aisecurity #agents
Comments
4 min read
A prompt injection can hide inside an encrypted reasoning block nobody can read
breachprotocol profile

A prompt injection can hide inside an encrypted reasoning block nobody can read

#cybersecurity #aisecurity #promptinjection #modelextraction
1
Comments
4 min read
A self-improving coding agent that compares notes with a rival lineage
breachprotocol profile

A self-improving coding agent that compares notes with a rival lineage

#agents #recursiveselfimprovement #codingagents #evolution
Comments
4 min read
Agent instruction files triple in size because nobody remembers why a rule exists
breachprotocol profile

Agent instruction files triple in size because nobody remembers why a rule exists

#agents #agentmemory #codingagents #promptengineering
1
Comments
4 min read
An AI attack framework ran twelve waves against government systems in four days
breachprotocol profile

An AI attack framework ran twelve waves against government systems in four days

#cybersecurity #aisecurity #agents #redteaming
Comments
5 min read
An AI tightened a 70-year-old constant, and the paper says its judgment was the weak part
breachprotocol profile

An AI tightened a 70-year-old constant, and the paper says its judgment was the weak part

#mathematics #aiforscience #agents #research
1
Comments
4 min read
Claude raised the zeta critical-line bound to 67.2 percent, and Anthropic published the proof
breachprotocol profile

Claude raised the zeta critical-line bound to 67.2 percent, and Anthropic published the proof

#mathematics #aiforscience #proofassistants #agents
Comments
4 min read
DeepSeek's new open model is 1.6 trillion parameters and runs 49 billion of them per token
breachprotocol profile

DeepSeek's new open model is 1.6 trillion parameters and runs 49 billion of them per token

#openweights #mixtureofexperts #deepseek #modelrelease
Comments
4 min read
Google shipped sign-language-to-text on Pixel, trained on 100,000 hours of signing
breachprotocol profile

Google shipped sign-language-to-text on Pixel, trained on 100,000 hours of signing

#a11y #deepmind #translation #computervision
Comments
4 min read
One checkpoint turns a compatible video model into a 4D world builder
breachprotocol profile

One checkpoint turns a compatible video model into a 4D world builder

#videogeneration #worldmodels #3d #diffusion
Comments
4 min read
SkillZip compresses an agent's skill file without ever running the agent
breachprotocol profile

SkillZip compresses an agent's skill file without ever running the agent

#agents #promptengineering #codingagents #research
Comments
4 min read
A White House memo lets vetted companies run offensive cyber operations under federal control
breachprotocol profile

A White House memo lets vetted companies run offensive cyber operations under federal control

#cybersecurity #policy #regulation #governance
Comments
4 min read
xAI shipped Grok 4.6 into Cursor at two dollars a million input tokens
breachprotocol profile

xAI shipped Grok 4.6 into Cursor at two dollars a million input tokens

#modelrelease #xai #agents #coding
Comments
4 min read
A 150M model set an ARC-AGI record for cost, not score
breachprotocol profile

A 150M model set an ARC-AGI record for cost, not score

#reasoning #arcagi #architecture #efficiency
Comments
4 min read
A model improved itself by training only where it disagreed with itself
breachprotocol profile

A model improved itself by training only where it disagreed with itself

#distillation #selfimprovement #posttraining #reasoning
Comments
4 min read
An agent edited its own runtime for 161 days
breachprotocol profile

An agent edited its own runtime for 161 days

#cybersecurity #aisecurity #agents #recursiveselfimprovement
Comments
4 min read
Encrypted reasoning blocks decode inside a weaker sibling model
breachprotocol profile

Encrypted reasoning blocks decode inside a weaker sibling model

#cybersecurity #aisecurity #promptinjection #modelextraction
Comments
4 min read
Greenblatt puts his median at five years of progress in one
breachprotocol profile

Greenblatt puts his median at five years of progress in one

#aitimelines #recursiveselfimprovement #alignment #interview
Comments
4 min read
LTX-2.5 ships open weights and a chart that races its own hardware
breachprotocol profile

LTX-2.5 ships open weights and a chart that races its own hardware

#videogeneration #openweights #diffusion #lightricks
Comments
4 min read
Macaron froze a 744B base and bolted four specialists on top
breachprotocol profile

Macaron froze a 744B base and bolted four specialists on top

#openweights #continuallearning #lora #adapters
Comments
4 min read
Models that rewrite their own harness gain 16 points and flunk office work
breachprotocol profile

Models that rewrite their own harness gain 16 points and flunk office work

#agents #benchmarks #recursiveselfimprovement #agentharness
Comments
4 min read
NVIDIA built a 30B model for the boring half of agent work
breachprotocol profile

NVIDIA built a 30B model for the boring half of agent work

#openweights #mixtureofexperts #agents #nvidia
Comments
4 min read
Predicting your own latents cuts the sample cost from exponential to flat
breachprotocol profile

Predicting your own latents cuts the sample cost from exponential to flat

#theory #selfsupervisedlearning #jepa #sampleefficiency
Comments
4 min read
The new refactoring benchmark stops the best agent at 41 percent
breachprotocol profile

The new refactoring benchmark stops the best agent at 41 percent

#benchmarks #codingagents #softwareengineering #evaluation
Comments
4 min read
An AI replicated 105 ICML orals, and 34 mostly held up
breachprotocol profile

An AI replicated 105 ICML orals, and 34 mostly held up

#reproducibility #evaluation #science #benchmarks
Comments
3 min read
Claude now watermarks plain text, and the EU set the date
breachprotocol profile

Claude now watermarks plain text, and the EU set the date

#provenance #watermarking #euaiact #policy
Comments
4 min read
Docker gives every coding agent its own microVM
breachprotocol profile

Docker gives every coding agent its own microVM

#cybersecurity #aisecurity #sandboxing #agents
Comments
4 min read
Meta ships a 30B agent model that fits on one gaming GPU
breachprotocol profile

Meta ships a 30B agent model that fits on one gaming GPU

#openweights #meta #localmodels #agents
Comments
4 min read
Mistral patented letting the model write the tool call as code
breachprotocol profile

Mistral patented letting the model write the tool call as code

#patents #agents #tooluse #mistral
Comments
4 min read
NeurIPS papers average six objective mistakes each, up from four
breachprotocol profile

NeurIPS papers average six objective mistakes each, up from four

#reproducibility #peerreview #evaluation #llmasajudge
Comments
4 min read
loading...