ついはじめ | Hajime Tsui
hajimetwi3.bsky.social
ついはじめ | Hajime Tsui
@hajimetwi3.bsky.social
Benevolent Influence Research (BIR) — Founder
Inventor of Post Engineering for AI
Researcher / Engineer
(AI, Cyber Security, Software, Infrastructure)

X: https://x.com/hajimetwi3
GitHub: https://hajimetwi3.github.io/hajimetwi3/
Pinned
My thread on X announcing the "Post Engineering for AI (PE4AI)" preprint just hit 10k views! If you have a moment, please check it out.

X Announcement:
x.com/hajimetwi3/s...

#Promptinjection #LLM

* Video generated with Sora
やあ。
エックスで障害が発生したようなので、ブルースカイを覗きにきました👀
January 23, 2026 at 1:55 PM
Reposted by ついはじめ | Hajime Tsui
I'm sharing a preprint titled "Meta Attacks for AI: Human and AI Reviewers Under Attack".

Explicit attacks that target the guardrails themselves remain underexplored.

In this work, I examine this problem and define the following concepts:

doi.org/10.5281/zeno...

#cybersecurity
#infosec
Meta Attacks for AI: Human and AI Reviewers Under Attack
Currently, guardrails are implemented in generative AI systems, such as chat-based AI, to detect malicious users and to prevent harmful outputs through automated filtering and content moderation in hu...
doi.org
January 22, 2026 at 2:43 AM
I'm sharing a preprint titled "Meta Attacks for AI: Human and AI Reviewers Under Attack".

Explicit attacks that target the guardrails themselves remain underexplored.

In this work, I examine this problem and define the following concepts:

doi.org/10.5281/zeno...

#cybersecurity
#infosec
Meta Attacks for AI: Human and AI Reviewers Under Attack
Currently, guardrails are implemented in generative AI systems, such as chat-based AI, to detect malicious users and to prevent harmful outputs through automated filtering and content moderation in hu...
doi.org
January 22, 2026 at 2:43 AM
My thread on X announcing the "Post Engineering for AI (PE4AI)" preprint just hit 10k views! If you have a moment, please check it out.

X Announcement:
x.com/hajimetwi3/s...

#Promptinjection #LLM

* Video generated with Sora
January 19, 2026 at 3:33 PM
PE4AIというプレプリントを発表しているので、ご紹介させてください。

「Post Engineering for AI: Benevolent Contextual Guidance for Debiasing Large Language Models」
doi.org/10.5281/zeno...

これは、善意でLLMの中立性等を底上げをするという日本発のアイデアで、
善意のPrompt injection、Context engineering、Data Poisoningの要素を組み合わせたものになっています。

#cybersecurity
#infosec
#生成AI
Post Engineering for AI: Benevolent Contextual Guidance for Debiasing Large Language Models
This paper proposes Post Engineering, a novel benevolent prompt-injection and contextual-influence technique in which neutrality-oriented guidelines are embedded at the end or elsewhere in publicly vi...
doi.org
January 17, 2026 at 12:08 AM
"つい"Bluesky始めました🎉
(某SNSが通信障害だったため…)

最近は、ついはじめ(Hajime Tsui)名義で生成AIやサイバーセキュリティ等の研究・発表を行っています。
こちらでもよろしくお願いいたします。

プレプリントやプロジェクト一覧を以下に記載しています。ご興味のある方はご覧いただけますと幸いです。
hajimetwi3.github.io/hajimetwi3/

#Bluesky
#JPTips
#cybersecurity
Hajime Tsui - Researcher / Engineer
hajimetwi3
hajimetwi3.github.io
January 16, 2026 at 9:07 PM