Umesh Wanve

AI/ML, Security Researcher, Reverse Engineer, Malware/Botnet/Exploit Analyst, Blogger, Bug hunter, Ethical Hacker & a lifetime learner.

[Virus Bulletin @virusbtn]

Symantec Threat Hunter Team reports GodDamn, the latest rebrand of Beast ransomware, itself a rebrand of Monster from the Hyadina developers. Operators used AnyDesk, NirSoft credential tools, & Microsoft-signed PoisonX driver to disable endpoint defences. https://t.co/hxCiUgpf4x

[Virus Bulletin @virusbtn]

Hybrid Analysis reports fake cryptocurrency wallets distributed as oversized Advanced Installer packages. The campaign impersonates wallets (e.g. Anchor, Zec, Iota, Onto, Dark, Stellar) using search-ranked attacker domains & valid code-signing certificates https://t.co/kAd3B7W2ns

[RedTeam Recipes | RTR @RedTeamRecipes]

We just dropped a comprehensive engineering guide on PowerShell Defense Evasion, detailing how to systematically dismantle these protections.

Read the full technical breakdown:

https://t.co/dfaBYXihRQ

#DefenseEvasion #RedTeaming #OffensiveSecurity #PowerShell

[Havox @Havox03]

DynLoader

A modular Windows loader focused on EDR evasion Built with indirect syscall (Tartarus Gate / Hell’s Gate), Manual PE parsing & Section mapping, API hashing & Dynamic resolution, fileless HTTP delivery (AES encrypted option), & more :)

https://t.co/N0wpp3yRMA

[📕「マルウエアの教科書」著者 | 吉川孝志 | 増補改訂版🌟発売中 @MalwareBibleJP]

マルウェア解析をAIエージェントに任せるとき、成否を分けるのは使うモデルの賢さではなく解析の「段取り」だった、という実測比較の記事が公開されています。題材は2011年にCCC(ドイツのハッカー集団)が公表した政府製スパイウェア「Staatstrojaner」の検体mfc42ul[.]dll。逆アセンブラをMCPで接続しただけの素のエージェントは、Skype傍受や永続化、プロキシ探索、C2通信といった見えやすい機能は素早く洗い出す一方、埋め込まれた暗号鍵や深い制御構造までは届きませんでした。そこに再利用可能な手順(スキル)と、解析の途中状態をディスクに残す仕組みを与えて初めて内部構造まで踏み込めた、とされています。

【要点の整理】

・素のエージェント(逆アセンブラをMCPでつなぐのみ・作業手順の指定なし)は、約15分・268回のツール呼び出しでドライバ投下やSkype傍受、プロキシ探索、C2の通信経路までは的確に洗い出す一方、暗号処理・埋め込み鍵・深い制御構造には届かず、コンテキストウィンドウをほぼ使い切ったとの結果。

・エージェントのモデルを替えても傾向は同じで、別モデルでは録音処理やVM検知などが数点増える程度。深部の再構成に効くのはモデルの差し替えより、明確な手順と、コンテキストが圧縮されても失われない作業状態の保持だという整理。

・改善版は段階を切った構造化ワークフロー(広い指標収集→分類→仮説立て→内部構造の地図化→深掘り計画→深掘り解析)で、統括・計画・報告の役割を分け、ディスク上のケースフォルダを外部記憶として使う設計。コンテキストが圧縮されても過去の成果を読み直して続行できる構成。

・同じ検体を構造化ワークフローで解析し直すと、10種類のコマンドを振り分けるディスパッチ用のvtable、ホストプロセスに応じた起動判定、C2のプロトコル層と暗号ルーチンまで復元。通信のワイヤ形式が独自マーカー(C3PO-r2d2-POE)＋4バイト長＋暗号化ペイロード、という深さに到達したとの報告。

・ただし完全な再現ではなく、ハードコードされたC2サーバーや埋め込みAES鍵の抽出には至らず、暗号方式もAES-128-ECBと報告したが実際はAES-256-ECBという取り違えも残存。実行のたびに結果が変わる非決定的な性質も明記され、出力は解析者による検証を前提とした出発点にとどまるという位置づけ。

・本件は同じ研究者による一連の素材の一部で、以前こちらでこのテーマのウェビナー(Binary-Cartography第2回)を紹介ポスト済み。今回はその文章版にあたり、別検体で素のエージェントと構造化パイプラインを具体的に比較した点が新しい部分。

エージェントが解析者を置き換えるのではなく、証拠集めや最初の構造把握といった機械的な部分を高速に肩代わりする一方、強力な難読化や仮想化、独自パッキングが絡むと途端に行き詰まるため、人の専門性はむしろ難読化解除のような領域に集中していく、という見立てで締めくくられています。

詳細は以下を参照：
https://t.co/xFk4PKMvfb

解析環境・パイプライン(GitHub)：
https://t.co/hmXPAbXkqK

(※可能な範囲でファクトチェックは実施済なものの、元記事の精度依存や速報・要約の性質上漏れもありうるため、正確な情報は一次情報を直接参照のこと)

[Min Choi @minchoi]

Less than 24 hours ago, Anthropic dropped Claude Fable 5.

Minds are blown. And people are already coming up with wild use cases.

10 examples:

[Neo Kim @systemdesignone]

If you're serious about AI engineering (in 2026), then learn these 13 concepts:

1 How Vector Database Works
→ https://t.co/FVxan8xHH3

2 How RAG Works
→ https://t.co/cGmunPTUlb

3 Design Personal Chat Assistant
→ https://t.co/nNWq3onTnW

4 LLM Concepts - A Deep Dive
→ https://t.co/5lCKxq2g4N

5 How to Design an AI Agent
→ https://t.co/JvnPd9773A

6 What is Reinforcement Learning
→ https://t.co/AVpl9j1oit

7 LLM Evals 101
→ https://t.co/nv3Ol8W53p

8 Context Engineering 101
→ https://t.co/OMkiZhkODL

9 AI Coding Workflow 101
→ https://t.co/paIf9ksIU9

10 Agentic Patterns, Simply Explained
→ https://t.co/8YdBBWvTj1

11 How AI Agents Work
→ https://t.co/tk3zkCjRvg

12 Multi-Agent Architectures, Clearly Explained
→ https://t.co/rS5QQS7Jln

13 How MCP Works
→ https://t.co/wgf8gHnnkn

What else should make this list?

===

👋 PS - Want my System Design Playbook (for Free)?

Join my newsletter with 200K+ software engineers now:
→ https://t.co/ByOFTtOihX

===

💾 Save now & repost to help others learn AI engineering.
👤 Follow @systemdesignone + turn on notifications.

[Radha Tripathi @Radha_AI]

Learn AI for free directly from top companies.

1 - Anthropic:
https://t.co/SIg8XJc2Xr

2 - Google:
https://t.co/b7oCmufjd3

3 - Meta:
https://t.co/EmgcvWclXp

4 - NVIDIA:
https://t.co/ki9yYO5x7M

5 - Microsoft:
https://t.co/BCoYgz87yr

6 - OpenAI:
https://t.co/FzNtuvavqx

7 - IBM:
https://t.co/uyvspbeF0r

8 - AWS:
https://t.co/uKkPz0evA2

9 - https://t.co/nStbl2HRBh:
https://t.co/eFCPvTYlbA

10 - Hugging Face:
https://t.co/K8DDuoaPxb

👇Comment "Learning" if you find this helpful.

Repost so others can take help.

Must bookmark for future reference.

[adam_cyber @Adam_Cyber]

It’s guaranteed to be a spicy event!

[Mandiant (part of Google Cloud) @Mandiant]

Secure the AI supply chain with #GoogleTIMondays. 🛡️

Use Code Insight to hunt malicious files & OpenClaw skills.

#CyberSecurity

[Security Harvester @secharvesterx]

[Tool] Rikune: A reverse-engineering MCP server that turns PE triage, Ghidra analysis, runtime evidence, and source-like reconstruction into reusable tools for any LLM
https://t.co/VJ77W1iPag

[adam_cyber @Adam_Cyber]

@CrowdStrike intel analysis of the recent Axios supply chain incident with attribution assessment:

https://t.co/lagCSjyVnq

[Meer | AI Tools & News @Meer_AIIT]

https://t.co/KQJ553jEIx

[adam_cyber @Adam_Cyber]

Incredible work by CrowdStrike Counter Adversary Operations and our broader team on this year’s report.

The trend line is clear: breakout times continue to accelerate. Defenders have less time than ever to detect, respond, and contain before impact.

AI is reshaping the battlefield. Adversaries are operationalizing it to scale social engineering, reconnaissance, and development — while AI adoption across enterprises is simultaneously expanding the attack surface.

Cloud intrusions are rising sharply, with identity compromise remaining the dominant initial access vector. We’re also seeing an unprecedented number of nation-state actors targeting cloud environments.

China-aligned activity remains globally expansive, with increased focus on logistics and critical infrastructure. DPRK operations continue to blend espionage and revenue generation, including crypto theft.

The data is deep. The trends are clear. The implications are urgent.

A must read!

https://t.co/OCQT24ylsB

[Akshay 🚀 @akshay_pachaar]

AI agents don't have a security problem.

They have an identity problem.

Here's what I mean:

You've deployed 47 agent workflows across your infrastructure.

At 3 am, an agent modifies a production database. You check your logs. 47 agents had access. They all used the same service account. You can't tell which one did it.

IBM's 2025 breach report found that 97% of organizations with AI-related breaches lacked proper AI access controls.

This happens because traditional security was built around human behavior: request access, get approved, complete the task, log out.

Agents don't work that way.

They run continuously for hours or days, making independent decisions at machine speed. They don't log out. They don't pause between tasks. They just keep acting.

Microsoft projects 1.3 billion agents in enterprises by 2028. Most will inherit long-lived API tokens and overprivileged service accounts.

The fix isn't better credentials. It's giving every agent its own identity.

Teleport is building an open-source framework for exactly this. I partnered with their team to show you how it works:

↳ Each agent gets its own cryptographic identity. No shared accounts, no static keys.
↳ Access decisions happen in real-time, scoped to what the agent needs right now.
↳ Every action traces back to a specific agent. Rogue behavior is detectable.
↳ Agents connect to databases and tools through the framework, not through scattered credentials.

Deploy agents safely. See exactly what each one does.

The framework is 100% open source. I have shared the link to the GitHub repo in the next tweet.

[Md Ismail Šojal 🕷️ @0x0SojalSec]

Open Source AI-powered Cyber Threat Intelligence & Hunting Platform.

- https://t.co/wG1iH17092

#infosec #cybersec #bugbountytips

[Ashot Gabrelyanov @gabrelyanov]

I’ve been struggling with the OpenClaw setup. Burned over $1,000 on tokens in just 3 days.

Then this guy saved me weeks of time and serious money.

Much respect.

If you’re setting up OpenClaw, this is the only post you need to read. 👇

[Umesh Wanve @Umesh_Wanve]

https://t.co/cE9fPca8vI

[adam_cyber @Adam_Cyber]

@crowdstrike is observing opportunistic eCrime actors l pushing Skrawl info stealer via fake OpenClaw skills on GitHub. One-line Terminal commands trick users into running it - steals Keychain, creds, crypto wallets.

See the blog we put out on open claw: https://t.co/0hXF7mXtaM

[Tech with Mak @techNmak]

What is RAG? What is Agentic RAG?

> Retrieval-Augmented Generation (RAG) <
----------------------------------------------

Retrieval-Augmented Generation (RAG) is an architecture that enhances a language model’s outputs by grounding them in external knowledge sources at inference time.

Instead of relying solely on parameters learned during training, RAG systems dynamically retrieve relevant information and inject it into the model’s context before generation.

=> Canonical RAG workflow
> A user submits a query.
> The query is embedded and matched against a pre-indexed corpus (commonly stored in a vector database).
> The top-K most relevant document chunks are retrieved.
> Retrieved context is appended to the original query.

A language model generates a response conditioned on this augmented input.

=> Primary objective
To reduce hallucinations and improve factual accuracy by grounding generation in verifiable, external context.

=> Key limitation
> Traditional RAG is a single-shot pipeline:
> No explicit reasoning or planning
> No validation of retrieved evidence

No iterative refinement if retrieval or generation is suboptimal

The system assumes the first retrieval and generation pass is sufficient, which often breaks down for complex, ambiguous, or multi-hop queries.

> Agentic RAG <
------------------

Agentic RAG extends standard RAG by introducing autonomous decision-making agents that can reason, plan, evaluate, and adapt across multiple steps.

Rather than a static retrieval → generation flow, Agentic RAG operates as a closed-loop, goal-driven system.

=> Core idea
Retrieval and generation are no longer treated as isolated steps, they become actions taken by agents in pursuit of a higher-level objective: producing a correct, complete, and useful answer.

=> Typical Agentic RAG Architecture
1./ Planning Agent
> Interprets the user’s intent
> Decomposes complex queries into sub-tasks
> Determines what information is required and from which sources

2./ Retrieval Agent
> Dynamically reformulates search queries
> Retrieves information from:
- Vector databases
- Structured databases
- APIs
- Tools or live data sources
> Can perform multi-hop retrieval when needed

3./ Generation Agent
> Synthesizes retrieved evidence into a coherent response
> Reasons across multiple sources
> Maintains traceability between claims and evidence

4./ Evaluation (Judge) Agent
> Critically reviews the generated output
> Checks for completeness, correctness, and alignment with the original query
> Decides whether to:
- Accept the answer
- Refine retrieval
- Re-plan and regenerate

This feedback loop can repeat until predefined quality criteria are met.