Raúl Ferrer

Raúl Ferrer

@raulferrer.dev

AI & Backend Architecture | Ex-Mobile Tech Lead | 27+ yrs in Tech Redesigning the future with agentic AI, multi-agent systems, and robust backends. 📘 Author @ Apress: shorturl.at/eqi6v ✍️ Medium: https://raulferrergarcia.com

952 followers1.3K following90 posts

Pinned

Raúl Ferrer
@raulferrer.dev

I’ve been quiet for some time, when I last spoke to you about Flutter and mobile development. After bringing that area to a mature, stable, and smoothly running state, the time has come to take the next natural step in my career. 🧵👇 #AgenticAI

Raúl Ferrer
@raulferrer.dev

Master AI Engineering from Scratch! 🧠 523 lessons & 20 phases 📐 Build every algorithm from raw math to code (Python, TS, Rust, Julia) ✨ 100% free & open source Start your interactive path today: aiengineeringfromscratch.com/index.html #AI #AIEngineer #Learn

aiengineeringfromscratch.comAI Engineering from Scratch523 lessons. 20 phases. Write the backprop, the tokenizer, the attention mechanism, and the agent loop by hand before any framework gets imported. Python, TypeScript, Rust, Julia.
Raúl Ferrer
@raulferrer.dev

No OpenAI IPO in 2026. Sam Altman says going public now would be "ill-advised" amid rising AI safety and alignment concerns across the frontier labs. #ai #openai fortune.com/2026/09/12/s...

fortune.comSam Altman confirms OpenAI won't go public this year, saying an IPO now would come at an 'ill-advised moment' given AI safety concerns | Fortune"I would say not 2026. Yeah, we got a lot of stuff to do, like meeting this moment of what is going to be required for safety and alignment, and how the industry and governments can work together."
View full thread
Raúl Ferrer
@raulferrer.dev

Were you already familiar with this report? What do you think about this transition toward agentic attacks? Let me know in the comments.

View full thread
Raúl Ferrer
@raulferrer.dev

>>> AI-Augmented Cyber Ops: Bad actors aren't just asking for code anymore—they're orchestrating multi-agent swarms for active recon, zero-day discovery, and self-adapting malware.

Raúl Ferrer
@raulferrer.dev

Some of the best Claude Skills repositories to supercharge your AI agent: - anthropics/skills - obra/superpowers - multica-ai/andrej-karpathy-skills - mattpocock/skills - nextlevelbuilder/ui-ux-pro-max-skill - JuliusBrussee/caveman - addyosmani/agent-skills - Leonxlnx/taste-skill #AI #claude

Raúl Ferrer reposted

Timothy Hughes
@timothyhughes.bsky.social

#TimTalk - AI Transformation Strategy: Fixing Your Unseen Operating Model with Liz Henderson buff.ly/u3el0as via @DLAIgnite #SocialSelling #DigitalSelling #Innovation #Technology #Tech #TechNews #AIAgents #AI #AgenticAI #GenAI #Leadership #Strategy #CIO #CTO #AI #BigData

Raúl Ferrer reposted

Bill
@sempf.infosec.exchange.ap.brid.gy

Well written piece by Brent Huston with the latest look at AI attacks. https://stateofsecurity.com/the-top-three-things-security-teams-need-to-understand-about-ai-driven-attacks/ #agenticai #cybercrime

stateofsecurity.comThe Top Three Things Security Teams Need to Understand About AI-Driven AttacksThe easiest mistake security teams can make about AI-driven attacks is to imagine a completely new species of adversary. That is not what most organizations are facing. The objectives remain familiar: steal credentials, gain access, persist, move laterally, collect data, commit fraud, extort the victim, or conduct espionage. What AI changes is the economics of the attack. It reduces the time, expertise, and attention required to move from one step to the next. That distinction matters because it tells defenders where to act. Google Threat Intelligence Group reported this week that it has observed adversaries moving beyond basic prompting into agentic workflows and AI-enabled automation. In one Q2 2026 case, a threat actor used an AI coding chatbot and agent instructions to help plan, build, and execute a mass credential-harvesting campaign in less than six hours. The framework automated scanning, troubleshooting, and IP rotation with limited human involvement. (Google Threat Intelligence Group, September 8, 2026) That does not mean every attacker has become an autonomous machine. It means the constraints that used to slow attackers down are weakening. Security teams need to understand three things. ## 1. AI Changes the Economics of an Attack Before It Changes the Objective AI is most useful to attackers when it compresses work. Reconnaissance can be summarized. Public vulnerability information can be translated into testable ideas. Phishing lures can be localized for a target’s language, role, industry, and current concerns. Malware can be debugged. Stolen data can be classified. Infrastructure can be configured. Failed steps can be revised without waiting for a specialist to return. This is important because many defensive assumptions quietly depend on attacker friction. We assume a small criminal group cannot personalize messages for thousands of targets. We assume an inexperienced operator will make obvious technical mistakes. We assume the time between vulnerability disclosure and reliable exploitation will give us room to patch. We assume the volume of stolen data will slow an attacker before the most valuable records are identified. AI weakens each of those assumptions. Microsoft Threat Intelligence reported in March that most observed malicious AI use still centered on producing text, code, or media. Attackers were using it to draft lures, translate content, summarize stolen data, generate or debug malware, and scaffold scripts or infrastructure. Human operators generally retained control over objectives, targeting, and deployment decisions. (Microsoft Threat Intelligence, March 6, 2026) OpenAI’s February threat report made a similar point: malicious actors typically combine AI with traditional tools, websites, social accounts, and other platforms rather than conducting an entire operation inside one model. (OpenAI, February 25, 2026) The first defensive lesson is therefore simple: > Do not look only for an AI artifact. Look for an attack chain whose cost and cycle time have changed. An AI-written email may not carry a reliable linguistic signature. AI-assisted malware may still use familiar persistence and command-and-control techniques. AI-summarized stolen data still has to leave the environment. The defender’s best opportunity may be the behavior around the content, not proof of how the content was produced. Security teams should ask: * Which attack steps used to require scarce expertise? * Which controls assume the attacker will be slow, noisy, or inconsistent? * Which defensive queues already move more slowly than the exposure they manage? * Where could an attacker use AI to test, fail, adapt, and try again at low cost? The answer will usually point to a handful of urgent bottlenecks: exposed systems, identity recovery, fraud verification, vulnerability triage, detection engineering, and containment approval. ## 2. Agentic Autonomy Is Real, but It Is Uneven The industry tends to make one of two errors. The first is dismissal: AI is only producing better phishing emails. The second is exaggeration: fully autonomous systems are already discovering unknown vulnerabilities, compromising enterprises, and completing every objective without human direction. The evidence supports a more disciplined position. Agentic systems are beginning to perform meaningful portions of real attack workflows. They can observe results, choose the next action, call tools, correct errors, and continue toward an objective. That is more consequential than generating a block of code in a chat window. At the same time, autonomy varies by actor, task, tooling, access, and model reliability. In its September report, Google described operational agentic credential-harvesting activity, but also said it had not yet observed threat actors deploy fully autonomous zero-day exploitation pipelines against targets in the wild. Its observations instead showed a gradual maturation: adversaries were using AI to accelerate n-day exploit development, refine payloads, and assemble multi-stage chains. (Google Threat Intelligence Group, September 8, 2026) Anthropic reached a related conclusion after mapping 832 accounts banned for malicious cyber activity between March 2025 and March 2026 to MITRE ATT&CK. Its analysis found AI use moving deeper into post-compromise activity. It also argued that the more durable signal of a dangerous actor is not simply the number of techniques used, but the scaffolding built around the model to chain attack stages together with less human input. (Anthropic, June 3, 2026) This changes threat modeling. Security teams should stop treating autonomy as a yes-or-no label. Instead, examine the attacker’s control loop: 1. What can the system observe? 2. Which decisions can it make? 3. Which tools or credentials can it use? 4. How does it determine whether an action succeeded? 5. What causes it to stop, escalate, or change tactics? The faster that loop runs, the less time defenders have between initial signal and material harm. That makes defensive latency a security control. Teams need high-confidence detections for identity changes, token use, privilege escalation, abnormal cloud activity, suspicious repository access, and unusual data movement. They also need preauthorized, bounded containment actions that can slow an attack without waiting for an improvised executive meeting. The objective is not to automate every response. Some actions can disrupt production, destroy evidence, or help the attacker understand what the defender knows. The objective is to decide in advance which actions can safely happen at machine speed, which require human judgment, and how quickly that judgment must arrive. ## 3. AI Is Both an Attacker Tool and an Attack Surface Most discussions about AI-driven attacks focus on what an adversary can do with a model. Security teams must also consider what an adversary can do to the organization’s AI systems. Models, prompts, agent instructions, source code, API credentials, cloud compute, vector stores, development tools, model gateways, and connected data sources now have value. They can be stolen, manipulated, poisoned, hijacked, or used as paths into other systems. Google reported Q2 2026 incidents involving theft of proprietary models, prompts, skills, code, and AI research, along with campaigns that targeted AI credentials and victim cloud resources for unauthorized model workloads. It also described malicious open-source packages and project files designed to influence AI coding assistants or interfere with LLM-based security scanning. (Google Threat Intelligence Group, September 8, 2026) This creates two connected threat models. The first is **AI for the attack** : the adversary uses AI to accelerate reconnaissance, social engineering, exploit development, credential harvesting, post-compromise work, or data analysis. The second is **AI as the attack path** : the adversary targets the AI supply chain, manipulates content consumed by an agent, steals model or service credentials, poisons instructions, or abuses the tools and permissions connected to an AI workflow. A security program that watches only for AI-generated phishing sees only the first few feet of the problem. The AI environment should be governed like privileged production infrastructure. At minimum, security teams should know: * Which models, agents, coding assistants, gateways, and AI services are in use. * Which identities and API credentials they use. * Which tools, repositories, data stores, and production systems they can reach. * Which untrusted content can influence their behavior. * Where prompts, instructions, skills, memory, and configuration are stored. * Whether tool calls and administrative changes are logged. * How an AI service or agent can be isolated, suspended, or reconstituted. The same principle applies to synthetic media and impersonation. In July, the FBI warned about scammers using AI-generated videos and spoofed sites to impersonate FBI personnel and revictimize people who had already reported fraud. The practical defense was independent verification through known official channels—not confidence that a person could visually identify every fake. (FBI Internet Crime Complaint Center, July 20, 2026) That is a useful operating rule for enterprises as well: > When a request can move money, change access, expose sensitive information, or alter production, verify the authority and the transaction through a path the request itself did not supply. Do not make perfect deepfake detection the control on which the business depends. ## Use the Three-Question Test For every significant incident, tabletop exercise, threat model, or intelligence review, ask three questions: ### Compression Which parts of the attack became faster, cheaper, more personalized, or easier to repeat because of AI? ### Autonomy Which observations, decisions, and actions could be chained together without waiting for a human operator? ### Exposure Which AI assets, credentials, tools, data sources, and trust relationships could become targets or execution paths? These questions are more useful than asking whether an attack is “AI-powered.” That label does not tell the incident commander what to contain. It does not tell the vulnerability team what to patch first. It does not tell the IAM team which recovery path to harden. It does not tell the board whether the organization’s risk has materially changed. Compression, autonomy, and exposure do. ## What Security Teams Should Do This Week Start small and make the work concrete. 1. **Update one attack scenario.** Take a familiar credential-theft, ransomware, fraud, or supply-chain scenario and identify where AI could compress the attacker’s work. 2. **Find one defensive delay that matters.** Measure the time from the first credible signal to the first action that meaningfully constrains the attacker. 3. **Inventory the AI control surface.** Identify production agents, coding assistants, model credentials, connected tools, repositories, and sensitive data paths. 4. **Instrument the control loop.** Log identity use, tool calls, configuration changes, unusual compute consumption, repository access, and consequential agent actions. 5. **Rehearse independent verification.** Test a high-consequence request involving a convincing executive, vendor, help-desk, or finance impersonation. None of this requires abandoning the security fundamentals. Phishing-resistant authentication still matters. Least privilege still matters. Vulnerability prioritization still matters. Segmentation, logging, egress control, secure software development, tested recovery, and practiced incident response still matter. AI does not repeal those controls. It changes the speed and scale at which their weaknesses can be discovered, combined, and exploited. The question security leaders should ask is no longer: > Was AI involved in this attack? The better questions are: > Where did AI compress the attack chain, what did it allow the attacker to automate, and which AI systems became part of the path? Teams that can answer those questions will be able to adapt without chasing every dramatic headline. Teams that cannot may discover that the attacker did not need a fundamentally new technique. It only needed to perform the old ones faster than the organization could respond. * * * ## More Information and Assistance MicroSolved, Inc. can help organizations: * Assess AI-enabled threat exposure and attack paths. * Threat-model AI agents, coding assistants, and connected workflows. * Review identity, API credential, data-access, and tool-use controls. * Test AI-accelerated fraud, credential theft, supply-chain, and incident-response scenarios. * Design practical detection, containment, and recovery improvements. Contact MicroSolved at info@microsolved.com or +1.614.351.1237. Relax. We’re on watch. _AI tools were used as a research assistant for this content, but human moderation and writing are also included._
Raúl Ferrer
@raulferrer.dev

I’ve been quiet for some time, when I last spoke to you about Flutter and mobile development. After bringing that area to a mature, stable, and smoothly running state, the time has come to take the next natural step in my career. 🧵👇 #AgenticAI

View full thread
Raúl Ferrer
@raulferrer.dev

At 50+😉, reinventing oneself is not a reaction; it is a strategic advantage. Bringing nearly three decades of design patterns, testing, and robust architecture to an AI ecosystem rife with fragile prototypes is where I see the real value today.

Raúl Ferrer
@raulferrer.dev

From now on, this space is evolving. I’ll be discussing backend architecture, autonomous agent orchestration, LLM evaluation, and how to take AI from a local script to production with engineering rigor. Is anyone else here designing or integrating multi-agent systems? Let’s connect. 🚀

Raúl Ferrer
@raulferrer.dev

I’ve been quiet for some time, when I last spoke to you about Flutter and mobile development. After bringing that area to a mature, stable, and smoothly running state, the time has come to take the next natural step in my career. 🧵👇 #AgenticAI

Raúl Ferrer
@raulferrer.dev

From the high-level frontend to the system engine: I have fully immersed myself in backend development and agentic AI architecture. Building interfaces gives you a nuanced understanding of the user; moving into the backend and multi-agent systems gives you control over the actual orchestration.

Raúl Ferrer
@raulferrer.dev

You've probably seen in a lot of applications how selecting a menu button causes the content of that menu to slide sideways on the screen. That element is the 𝗱𝗿𝗮𝘄𝗲𝗿. In the following article I explain how you can make your own drawer in 𝗙𝗹𝘂𝘁𝘁𝗲𝗿. levelup.gitconnected.com/how-to-build...

levelup.gitconnected.comHow to Build a Custom Drawer in FlutterLet’s explore how to create an adaptive Drawer with both left and right Drawer views in Flutter.
Raúl Ferrer
@raulferrer.dev

🚀 Ready to go from developer to leader? I’ve launched “Code, Lead, Grow”, a newsletter packed with: ✅ Leadership lessons. ✅ Actionable strategies for devs. ✅ Tools & resources. Don’t just code—lead. codeleadgrow.substack.com

codeleadgrow.substack.comCode, Lead, Grow | Raúl Ferrer | SubstackThe newsletter will position you as a thought leader, offering a mix of personal stories, practical advice, and motivational content to help readers excel in their technical and leadership roles. Clic...

Raúl Ferrer reposted

Vasil
@iamvasil.bsky.social

Developing your soft skills is Gold. After 15 years of management experience, I still find things that I could have said or done differently and hearing more about it helps!

Raúl Ferrer@raulferrer.dev

Soft skills like emotional intelligence, adaptability, and storytelling are more than just an added bonus—they're your secret weapon for 2025. Explore the essential soft skills you need to learn and how to excel at them: medium.com/illumination...

medium.comThe Soft Skills You’ll Need to Thrive in 2025Master these 5 Soft Skills in 2025 to elevate your career and personal growth

Raúl Ferrer reposted

Bluesky Safety
@safety.bsky.app

1/ With more users joining Bluesky, we know how important it is to identify which accounts are real. Users deserve confidence that the accounts they interact with are authentic. Sharing some details about our impersonation policy 🧵

Raúl Ferrer reposted

Marques Brownlee
@mkbhd.com

Annual shoutout to the Keepa Amazon price tracker plugin, so you can check to see if every Black Friday sale is actually a sale, or just marked as one so you think you're getting a deal

Raúl Ferrer reposted

Bluesky
@bsky.app

📢 App Version 1.95 is rolling out now (1/6) Lots of bugfixes and performance improvements in this one!

Raúl Ferrer
@raulferrer.dev

Huawei unveiled HarmonyOS Next, claiming it’s free of Android code, alongside the Mate 70 and foldable Mate X6. The Mate 70, launching Dec. 4 in China, highlights Huawei’s shift to in-house chips and software despite US trade restrictions. www.engadget.com/mobile/smart...

engadget.comHuawei launches its Android-free OSWith launch of its Mate 70 and foldable Mate X6 smartphones, Huawei unveiled its latest mobile OS that no longer uses any Android open-source code.
Raúl Ferrer
@raulferrer.dev

#Kotlin 2.1.0 is out and introduces new features like guard conditions, non-local break/continue, and multi-dollar string interpolation. It includes updates to the K2 compiler, Kotlin Multiplatform, Kotlin/Native, Kotlin/Wasm, and improved Gradle compatibility. blog.jetbrains.com/kotlin/2024/...

blog.jetbrains.comKotlin 2.1.0 Released | The Kotlin BlogThe Kotlin 2.1.0 release is out! This version introduces new language features, tooling updates, performance improvements, and important fixes. Learn more!

Raúl Ferrer reposted

Jan Armbrust 
@n0rthk1n9.eurosky.social

Far too many apps ignore this, which is frustrating when filling out contact forms or anything that involves data that could be automatically filled. #buildinpublic #indiedev #swift #swiftui #iOS #dev #iosdev

SwiftNuggets: 🍗 { var nuggets: [💡] }@swiftnuggets.bsky.social

Swift Nugget 3 🍗: The .textContentType modifier controls which autofill hint a user gets displayed above the keyboard of a SwiftUI TextField💡Add it to fill out address data for example at checkout ✅ Try it out: github.com/n0rthk1n9/Sw... #buildinpublic #indiedev #swift #swiftui #iOS #dev #iosdev

Raúl Ferrer
@raulferrer.dev

I'm spending this weekend coming up with concepts for Medium posts in the future. A combination of themes related to personal development and coding. Keeping things easy and allowing the imagination to run wild. Which subjects are you interested in reading about?