Chevy Phillip (@chevonphillip) header
Chevy Phillip (@chevonphillip) profile picture

AppSec engineer securing AI agents in public. Hands-on security results and research at https://teapawt.net/t.co/0YITVAs7Vq. Every number is real.

New York
Joined May 2019
Photo by Chevy Phillip (@chevonphillip)Photo by Chevy Phillip (@chevonphillip)Photo by Chevy Phillip (@chevonphillip)Photo by Chevy Phillip (@chevonphillip)Photo by Chevy Phillip (@chevonphillip)Photo by Chevy Phillip (@chevonphillip)Photo by Chevy Phillip (@chevonphillip)Photo by Chevy Phillip (@chevonphillip)Photo by Chevy Phillip (@chevonphillip)Photo by Chevy Phillip (@chevonphillip)
Chevy Phillip retweeted
Save this to edit your videos with Gemini. #ai #prompt #chatgpt
8
17
124
7,675
Higher attack success in RL red-teaming doesn't mean a stronger injection defense. PIPETTE (arXiv 2611.24362, @nasokan et al.) rewards new attack strategies, then trains the defender on-policy. AgentDojo static ASR on Qwen3-4B: 33.29% to 0.48%. Not zero. chevyphillip.tech/blog/pipet…
1
1
6
257
Chevy Phillip retweeted
🤯 Kimi K3.1 Pro Abliterated is now 40% OFF $1.5 in / $0.15 cache / $3 output
15
4
1
216
6,548
Chevy Phillip retweeted
You finally close a new client. Then you spend the rest of the day setting them up tasks in ClickUp. Now you can just give Polar Browser the details and it handles the setup for you:
8
10
1
117
5,240
The @GoogleDeepMind team and @GeminiApp been cooking. Outstanding !
Introducing Gemini 4 Neon! Pushing the frontier across domains and punching above its model class in a meaningful way. To me, Neon feels like the right balance of intelligence, speed, and usability. Available now across @GoogleAIStudio, the API, and more!
133
Introducing Gemini 4 Neon! Pushing the frontier across domains and punching above its model class in a meaningful way. To me, Neon feels like the right balance of intelligence, speed, and usability. Available now across @GoogleAIStudio, the API, and more!
791
877
642
13,198
1,431,406
Chevy Phillip retweeted
More capacity has been added for Kimi K3.1 Pro Abliterated TPS should be way better now although prefill speed is still being worked on
3
4
1
129
7,187
Try it out today! Earn some bonus credits. @refuselessai refuseless.com/?ref=jqwaatlo
⚡️ Introducing Kimi K3.1 Pro Abliterated, the most powerful open model for cyber and coding. • Native MXFP4 • 1M context • Zero data retention $2.5 input / $0.25 cache / $5 output
1
120
Chevy Phillip retweeted
We are tracking a breach of a 3rd-party JS host impacting hundreds of high value websites. This is staggering.
266
1,799
312
12,799
2,916,783
Agents that edit their own prompts, memory, tools can adopt updates that quietly make them worse. arXiv 2611.16378 gates each self-update on fresh evidence: 0 of 3,000 calibrated campaigns took a harmful update. Uncalibrated, 99.9% did. Simulation only. arxiv.org/abs/2611.16378?utm…
3
119
Are meta glasses any good? What’s your best use-case?
146
Chevy Phillip retweeted
Learn harness engineering with a project-based course covering environment, state management, verification, and control mechanisms for AI coding agents. github.com/walkinglabs/learn…
11
120
910
42,690
Chevy Phillip retweeted
Abliterated DeepSeek V4.1 Flash 1017 is again FREE for new users for a limited time. • 1M context window • Native FP8 • Zero data retention Come try it on refuseless.com, just sign up and create your api key! Limited capacity on free tier, first come first serve.
14
9
2
338
14,687
FYI…Java is still a 🐐 of a language.
149
I built HexBrain: every publicly disclosed HackerOne report on one interactive map. 15,000+ reports by vulnerability class, weakness and program Full write-ups: how each bug was actually found 405,000+ CVEs, exploited ones in red Bypass and payload libraries It's free and still early, so tell me what breaks or what's missing. hexbrainai.com #bugbounty #infosec #bugbountytips
76
187
12
1,056
46,498
Go is so underrated, success stories like these are the reasons why it’s my go-to for API development. Rust, I’m sure, would’ve worked here as well, but at what cost? Congrats @resend
We rebuilt our email API to Go. It took 7 weeks and 140 PRs, and we've been running in production for weeks at roughly half the median latency and one seventh the memory usage per task. Here's what it took to move our most critical API to Go 👇
135
Adversarial training wastes steps on "pseudo-attacks" that raise loss but never flip the prediction. New paper (arXiv 2611.14046) pushes inputs along the top Fisher eigenvector instead: 72.50% AutoAttack accuracy on CIFAR-10 ResNet-50. Weaker on ViT. arxiv.org/abs/2611.14046?utm…
1
137
Introducing Gemini 4 Hydrogen, an extremely capable and fast model built to bring Gemini 4 to everyone! Available to all tiers of Google AI Studio, Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro + Ultra subscribers in Google Antigravity.
807
832
462
11,546
2,921,783