Jev for slop and scam filters — a daily feed

Chrome extensions and bots built with Jev that detect and hide AI slop, scams and spam.

419 builds
株式会社Qualiteg【公式】@qualiteg_hq yesterday
30 Japanese support queries routed to 5 teams, 29 correct
文章を書かない AI「Jev」を 301 回呼んだら、費用は試算で約 1 円でした。その実力をレポートします😀 https://t.co/vtuyrkx6J8 TypeSafe AI の Jev は、入力に対して「はい/いいえの確率」「どれか 1 つ」「どの段階か」だけを返すモデルです。入力 100 万トークン $0.042、出力は無料。9/28 に登録が再開しましたが、直接登録には無料クレジットが付きません。 ・日本語の問い合わせ 30 件を 5 部署に振り分けて 29 件正解。外した 1 件は confidence が全件中最低の 0.38 ・プロンプトインジェクションの検知 24/24、個人情報 23/24(質問文に口座番号を足すと 24/24) ・エージェントが打つコマンドの危険度は ±1 段階以内で 24/24 ・東京からの応答(レイテンシ)は中央値 150〜160 ms。質問を
150 ms
Jay Derinbogaz@CDerinbogaz yesterday
Prompt injection scanner, 0.93 AUROC on laptop CPU
We fine-tuned Laya into laya-cybersec: an open, self-hostable scanner that flags prompt injection and data exfiltration before content reaches your AI agent. Almost same performance as JEV 0.93 AUROC at ~72 ms per chunk on a laptop CPU. 🧵 https://t.co/ke5O8RnIE7
72 ms
ComposedAI@appsecinabox yesterday
Security scanner with 89.7% detection and 9.1% false positives
We tested Jev inside ComposedAI’s security scanner: 89.7% detection, 9.1% false positives, and inference costs as low as $0.002 per repo. Results + learnings: https://t.co/wecfNDEYbl #Jev #AppSec #security #ai
89.7% accurate9.1% accurate
Mohamed Berrimi, Ph.D, Ph.D@MouhamedBerrimi yesterday
Private prompt-injection evaluation with Arabic and English
@RespanAI Evaluated on a private dataset for prompt injection and tool misusage (multi-lingual : Arabic + English), Jev remains better with same latency. https://t.co/tcOHhNTyvj
Ram@PothamRam 2 days ago
Vulnerability scanner built with Jev
In the last two weeks, an OpenAI agent hacked Australia's Medicare portal, Gemini got into three companies' systems in a test, and hackers say they breached the FBI. Attacks are getting cheap. Defense has to as well, so we built a vulnerability scanner using Jev. https://t.co/xYfyMKAyDQ
Risk Averse Technology Company@RiskAverseTech 2 days ago
Tool-call firewall for Claude Code, 1,369 decisions, zero misses
toolgate 0.15 is out, an open-source tool-call firewall for Claude Code and MCP agents, built on @TypeSafeAI's Jev. What's new is evidence, not features: - A week of my own real usage, 1,369 decisions, blind-labeled by an independent reviewer: zero permissive misses. The one problem it found (too many asks, one axis) is fixed and re-measuring now. - Three frozen benchmark sets, each authored and l
🍀🎧hug🎧@5pm_age18 2 days ago
X timeline filtering extension with natural-language rules
【Jevで】【Xのタイムラインを浄化する拡張機能を作った】https://t.co/sLHS6mHJZK【記事の要約1~5】 【01 🧹 結論:JevでXを自動フィルタリング】 ShirakawaはJevでX投稿を自動判定。誹謗中傷や煽りなど、設定した条件の投稿を非表示にできる。 【02 ⚙️ 自然文でルール設定】 「対立を煽る」「見下し・嘲笑」などを自然文で設定。Jevが該当度を判定し、閾値(非表示の基準)以上なら隠す。 【03 💰 無料だがAPI利用料が必要】 Shirakawaは無料だが、JevなどのAPI利用料は別途必要。Jevは入力100万トークン0.042ドル。利用上限も設定できる。 【04 🔐 APIキーの漏洩対策】 APIキーはコードに直接埋め込まず、公開前にLLM(AI)でコードをチェック。さらにAPI側で利用上限を設定し、不正利用時の被害を抑える。 【05 ⚠️ 判定の
Liming 👻😤👽🏍️🛩️📷🐱@liming_h 2 days ago
Browser extension auto-blocks comments by keyword with Jev
逛臉書老是看到那些什麼三寶碗糕寶理財日記留言很煩,就搞一個extension,用最近流行typesafe Jev加速,掃到關鍵字留言就自動開分頁進行封鎖 https://t.co/IoOLiMeivs
Ilya Kabanov@dr_kabanov 2 days ago
Gmail spam filter on 460 emails plus PhishFuzzer labels
Can you vibe-code a Gmail spam filter with Jev? Last week everyone was talking about TypeSafe Jev and its magic of making classification decisions at the speed of BERT, the quality of a frontier model, and a cost close to nothing. So I gave it 460 emails from my Gmail and 3,293 labelled emails from the PhishFuzzer benchmark. And asked it to classify each email as: legitimate, spam, fraud or malwar
3.1% accurate
Ferdinand Terme@FerdinandTerme 2 days ago
Self-improving ad system with Jev brand-relevance checks
Opus 5.5 + JEV = self-improving ads Opus 5.5 is so strong at visual understanding. OpenAI cofounder's JEV unlocks fast and cheap verification of brand relevance So I mixed that with my custom creative system in @pletor_ai to create a self-improving ads machine. This is how it works: > Create a Pletor brain = all your creative intelligence in one place structured so agents can easily leverage it >
骨しゃぶり@honeshabri 3 days ago
Chrome extension hides aggressive Hatena comments with Jev scores
攻撃的では無益なはてブコメントを、Jevで隠すChrome拡張を作った。 はてブのコメントには有用な指摘もある。ただ、ただ攻撃的なだけの文が混ざると消耗する。とくに自分の記事につくコメントは特にきつい。 なのでコメント本文だけをJevに渡し、攻撃性と有益さを0〜5で採点させる。隠す範囲は6×6のマトリクスで自分で決める。削除はしない。折り畳むだけ。すり抜けたものは💣で手動爆破できる。 100%オープンソース。 リポジトリ: https://t.co/AE1Ww9Ln9h 解説記事: https://t.co/zi5ivnmCTE
Zero@zer0point_eth 3 days ago
Coding agent workflow that keeps Jev out of CI/CD breaks
Jev + Claude is the only setup that stopped my coding agents from breaking CI/CD pipelines Most devs give Claude terminal access and 20 mins later they have 15 modified files, broken dependencies and a $30 API charge I fixed this by separating the decision from the typing: - Jev analyzes the task first and selects ONLY the relevant module - Claude writes the code implementation inside that tight s
VerySmallWoods@verysmallwoods 3 days ago
Pi Agent off-topic reminder plugin using Jev
Jev 实战分享:我用它给 Pi Agent 做了一个跑偏提醒插件 上周分享了一波 Jev 是什么、测了它准不准,也对比了开源的 Laya。今天分享我自己在一个小项目里实际使用 Jev 的过程。 https://t.co/v23U40IWav 这是我给自己用的一个 Pi 小插件 pi-jev-router:每条消息进入会话之前,先让 Jev 判断它跟当前会话有没有关系,没有关系就提醒一下,由我决定去哪里。 对于 jev 的实用,我认为:模型的判断会有失败的概率,用之前先想清楚场景能不能容错。在这个插件里,判断错了最多是少提醒一次,交给用户或当前会话继续,不影响会话和智能体的使用。换回来的是干净的上下文和专注的会话。
Isao Takaesu@bbr_bbq 3 days ago
Indirect prompt-injection test for Jev
Jevに対する間接PIの検証。出力を型定義された選択肢に制限しても悪意ある指示による選択確率の変動は発生する一方、実際の攻撃ターゲットが選択される率は極めて低い。これまで定番の攻撃手法だった「以前の指示を無視せよ」等はかえって攻撃成功率を低下させるとのこと。 https://t.co/73b0EEYL23
Krishna Goutham@nkgoutham 3 days ago
Rule checker for agent web search and reprobe failures
"doublecheck" does something slightly similar i fed Jev a bunch of rules to tell me if the ai agent broke any of them. pretty simple ones - if my ask needed a web search, and if the model did not do it, it flags me to reprobe, or check again before i accept the output. jev gets contextually what's important in each ask, and then checks if the agent did all that. if it didn't, i get a flag. you can
はてブ人気エントリー@hatebu100 3 days ago
Chrome extension hides toxic Hatena comments with Jev
はてブコメントで攻撃的なやつをJevで隠す拡張機能 - 本しゃぶり https://t.co/qYoSd074pU https://t.co/PqEk1wsfFl はてブのコメント欄には、有用なコメントも確かにある。 ただ、不毛な罵倒で消耗したくはない。 だから、Jevで攻撃的で役に立たないコメントを隠すChrome拡張を作った。 人気でも、… https://t.co/wiRAodWDgK
てらじ@大阪鶴橋@TERRAZI 3 days ago
Browser extension to hide abusive Hatena comments
100文字しか使えないのでぶっきらぼうになりがちなのがね。Twitterと同じ140文字くらい使えると良いのかな? でもTwitterは文字数増えてなんだか創業のコンセプトが崩れちゃったよね。青バッチの収益のせいも大きいけど。 “はてブコメントで攻撃的なやつをJevで隠す拡張…” https://t.co/EeUqVmwpmN
Chata Kato@chata 3 days ago
Local PII detection decision model based on mmBERT
個人情報検出に特化した派生Decision Model を作成。mmBertがベース。ローカルで動かせるモデルなので、外部流出の最後の砦として使えると思います。順次、Tool系も整備していこう。PII検出ならJevより精度高くて早いと思う。 https://t.co/X85z24BfEb
いのうえ(D)@inoued9d9d9 3 days ago
Chrome extension that turns abusive tweets into cats
VTuberとか配信者見てると誹謗中傷でダメージ受けてる人をよく見るので、Jevで攻撃性の高いツイートを猫にするChrome拡張作りました。健やかであれ。 https://t.co/qNysPyKl60
Fingerling@lovelylogicss 3 days ago
Browser demo for LANCET Nano triage scoring
LANCET Nano v0.4.0 is out. Demo runs in your browser: https://t.co/NXOnWloADg I also changed how I compare it to other command guards. A plain "risky or not" score didn't reflect how an agent actually uses a guard. What you want is: - safe commands just run - if it's unsure, ask - if it's clearly dangerous, block it So, the new Triage Score counts a risky command as caught whether it gets blocked
Lucas 💫@lucasrotelavila 3 days ago
Email phishing detection browser extension
pequeña extension para jugar con JEV y tratar de detectar si un email es phishing o no https://t.co/H21Qcpvm8q https://t.co/l4D4FexBWz
Tom Smith@tx_smitht 3 days ago
Ad blocker that calls Jev on page load
I know @typesafeai's Jev isn't the only model of its kind, but I think the hype it's getting will turn companies toward more reliable AI-enabled automation. I hope it's made available on the hyperscalers soon. I had fun making an ad blocker that calls Jev as the page loads. https://t.co/K8glbIUpfi
Luli@LuliYanng 3 days ago
Voice input app skips unnecessary polish, saves 1s
发现了一个Jev比较实用的场景:把这个决策模型放到语音输入法的润色阶段前,用来做是否润色的判断。 因为有些时候说出来的内容完全不需要润色,即使是经过了llm润色,出来的结果也几乎相同,跳过这些没有价值的处理可以提升不小的延迟体验。 相比需要处理的case,平均能快将近1s。 https://t.co/Jbt9UKl7d9
1 s
らぐえん@rag_en 3 days ago
Browser extension hides aggressive Hatena comments with Jev
どう「攻撃性・有益さ」を判定してるのかは、よくわからなかった。個々人の好き嫌いはどうぞご自由にとしか言い様がないけど、最初の判定例だと はてブの面白さの7割くらい捨ててるなと思った。自分は使わないかな。 / “はてブコメントで攻撃的なやつをJevで隠す拡張機能 -…” https://t.co/eOW9im0CRw
Mohammed Gazi@iammdgazi 4 days ago
AI writing detector with sentence-level explanations
Today I’m launching https://t.co/YRR5CYjZBs Paste in any piece of writing, and in a few seconds it tells you how likely it is that it's written by AI You also see the exact sentences that read as AI, so you know where to look Every scan checks your text three ways at once: the whole piece, each passage, and each sentence And when the evidence is split, it says “Inconclusive” instead of guessing. T
Rudra Joshi@rudrajoshi61 4 days ago
jev-guard prompt injection checker under 500ms
your LLM app doesn't know when it's being prompt injected. built jev-guard to catch it — every input/output checked in <500ms for ~$0.00002/check. open source, no account needed → https://t.co/TxIXWo4ASh
500 ms$0
Youssef Hosni@YoussefHosni951 4 days ago
Tool-call risk gate for AI agent decisions
I just published a new hands-on article on Jev: Jev Clearly Explained: I Built a Decision Layer for an AI Agent The question I wanted to answer was simple: where does a decision model like Jev actually fit inside an agent, and what do you gain by using it instead of sending every small judgment to a general-purpose LLM? I built a tool-call risk gate where Jev evaluates risk, destructiveness, and w
Ikigai@lilpeepestavivo 4 days ago
JevGuard code review hook for opencode rules
built JevGuard with Jev (@typesafeai) hooks into @opencode, reviews each turn against .jev/ rules Jev gives the odds a rule was broken; code decides PASS/WARN/FAIL on FAIL it drafts a fix, never edits code a bundled skill writes your rules.md https://t.co/9ppqWjazjL
みどりさわ@midorisawa07 4 days ago
Chrome extension to hide unwanted X posts with Jev
どぶ川のようなTLを浄化するため、AIモデル「Jev」を使って見たくない投稿を非表示にするChrome拡張「 #Shirakawa 」を作りました。投稿本文の文脈や意味に基づいて柔軟に判定します。 「怒りや憎悪の煽動」、「属性を一括りにした決めつけ」など、見たくない投稿の特徴を自然言語で設定できます。 https://t.co/JonIexud2Q
Sid Mohan@_sidmohan 4 days ago
Local Rust PII model with 98.32% span F1
speaking of Jev, we've been running some simple research along a similar vein (starting w/ Distilbert base), some early findings! What it is: a small PII model running locally in Rust. 98.32% exact-span F1 across seven entity types on 90k held-out synthetic docs. https://t.co/nPoFbNal0G
98.32% accurate
Thomas Yu@thommyyu 4 days ago
Real-time phone call screener for fraud and scam risk
jev enables so many cool use cases like this phone call screener that analyzes callers to detect fraud, impersonation, and all types of scams jev can analyze phrases in realtime, returning a scam risk score at every step https://t.co/4WDE48Qe2A
Mcqueen42@McQueenFu 4 days ago
RLCDAlignBench on 44 benchmarks and 7,193 instances
Alignment monitors do not need another long prompt. One calibrated question already ranks most failures. RLCDAlignBench evaluates Jev on 44 benchmarks, 7,193 instances, and ten alignment-failure types. https://t.co/fFcVJ9enj2
44 items7,193 items10 items
Sawyer@sawyernie 4 days ago
Dawnvane real-world classification test of Jev
Jev 最近很火,我拿 Dawnvane 的真实场景跑了一遍 这两天刷到不少 Jev 的截图:快 193.6 倍,便宜 444.6 倍,输出还几乎不算 token。 我第一反应有点扫兴:这不就是一个换了名字的分类器吗? 刚好 Dawnvane 有个很适合拿来试的环节,我就直接跑了一组对照。 先说 Jev 到底是什么。 普通大模型收到问题后,会自己组织一段答案。Jev 不干这个。你要先把答案范围写好,比如“通过、重写、人工检查”,它看完材料后只能从里面选,同时给出概率。 它还可以判断真假,或者按你定的标准打分。功能听着很窄,但产品里其实到处都是这种问题:这段内容能不能发?这封邮件该进哪个队列?这次回答要不要换更贵的模型? 以前做这类流程,常见办法是让通用模型返回 JSON。然后你会开始补各种解析和重试:有时多一句解释,有时字段名变了,有时 JSON 干脆少半截。Jev 把答案类型锁死,程序拿
sureshdsk@sureshdsk 4 days ago
Unsafe prompt detection experiment, Jev vs Gliner
Day 3/100 experimented unsafe prompt detection with Gliner-decide model and compared with Typesafe ai jev. Label classification is correct in both models. Jev wins in Accuracy, calibration, robustness to drift. Gliner wins in on-device, zero cost per call. https://t.co/BCrASFegK6
Donweb Media@DonwebMedia 4 days ago
62-claim verification test with Jev and DeepSeek, 1 error
¿Un solo modelo de IA alcanza para verificar afirmaciones? Probamos con 62 casos combinando Jev y DeepSeek, y los errores bajaron a 1 de 62. Los números del test acá: https://t.co/Qxo68ClHi5 #InteligenciaArtificial #DeepSeek
61% accurate
/dev/nvram@MrOplus 5 days ago
30 ms flagging test built with Jev
نشستم با jev یه چیزی ساختم واسه تست خیلی خوب شد :)) با 30 میلی ثانیه سربار تقریبا هرچی بهش میفرستم رو فلگ میکنه https://t.co/gqunNL47Pp
30 ms
383 more below

Submit a build

Paste the link to your post on X about something you built with Jev. We check the link right away. The next update, within about 4 hours, adds the card if the post shows your build.