AI security is still early and may not be stressed until some point in 2027. "We are still on the very early part of the lower part of the bell curve with AI security. I don’t think we are long before we turn that corner though. But this is also dependent on the vendor. Fortinet is starting to put AI security solutions like prompt injection detection and AI proxy gateways in their products, and Palo Alto is trying to sell products for that like Prisma AIRs. I’d expect this obviously to be a very focused part of the 2027 go-to-market motion as it continues to become a more important part of the discussion within sales cycles. However, AI security technology is still going to be adopted at the largest companies first as the price is high and they can afford, but the more important reason for this is that these solutions are still largely untested and you will find more of the problems upfront." We need to be secure! $PANW $FTNT $CRWD

The biggest takeaway from Google/Mandiant’s AVDH isn’t “AI can find vulnerabilities.” It’s that the model is no longer the moat — the harness is. Threat modeling → entry-point discovery → context enrichment → hypothesis generation → multi-agent validation → human PoC. The next race in AI security won’t just be about who has the best model, but who can encode the best security methodology around it.

AI agents are everywhere at @Uber. It’s great to see, but the thing that keeps me up at night is how we are going to secure them. This is something that I have been thinking about for a while. Today, our agents run 50,000+ sessions per day across thousands of endpoints. And this isn't just engineering anymore. Employees across the company use agents that read code, run commands, call internal tools, analyze data, and act on real systems. That scale forced us to confront an important question: How do you secure agents when your security tools can't even see them? Traditional Endpoint Detection & Response (EDR) sees the file write, but not the prompt that triggered it. It sees the network call, but not the agent's reasoning. The intent, the thing that separates malicious from benign, is invisible. So we built Agentic Detection and Response (ADR): • Capture the full causal chain: prompt → reasoning → tool call → outcome, across Cursor, Claude Code, Codex, and every agent our employees use. • Triage cheaply: a fast, high-recall first pass handles the flood of benign sessions. • Reason deeply: only suspicious events get expensive LLM analysis, enriched with source code, threat intel, and policy context. • Red-team continuously: an offline explorer evolves hard attack variants before attackers find them. After 10+ months in production, the results speak for themselves: • Hundreds of credential exposures detected across 26 categories. • Shift-left prevention blocking secrets at 97.2% precision, before they ever leave the laptop. • Zero false positives on our enterprise benchmark, with 2-4x the F1 score of state-of-the-art baselines. • Every attack detected on AgentDojo, the public prompt injection benchmark. Just as valuable as the detections are the lessons from running this in production: • The workflow is the unit of security, not the individual tool call. Attacks hide in causally-linked chains that look benign step by step. • Credential leakage is a far more common operational issue than prompt injection. • Approval fatigue is real: when users approve 50+ actions per session, human oversight becomes a rubber stamp. You can't secure agents you can't observe. And nobody can solve this alone. That is why we recently joined the Open Secure AI Alliance (OSA), and why today we're taking the next step: open-sourcing ADR. The release includes the ADR Sensor, the detection framework, and ADR-Bench, the first enterprise agentic AI security benchmark: 302 tasks derived from real production telemetry and full coverage of all 17 attack techniques across 5 tactics, so the community can rigorously evaluate their own defenses. Code: Paper: The future of AI security won't be built behind closed doors. Excited to see what the community builds on it, and what we all learn together! @UberEng

It seems every modern security researcher is building its own AI Security solution right now. Low false positives rate is prioritised, because less triage means more efficiency. Still, making your tool's 'judging' phase too rigorous can lead to dropping valid severe vulnerabilities. I'd recommend to prioritise being able to find deeper, multi-level vulnerabilities, instead of optimising for 0 triage time. Doesn't matter - human or AI finding - validation is always required anyway. And of course, don't do what everyone else is doing. This is how you fail. Be different to succeed🫡

🤖 Kimi-K3 & GPT-5.6 Are Now This Powerful — Can Anyone Make Money Finding Bugs? Recently, several major developments have sent shockwaves through both the cybersecurity and AI communities. First, Kimi-K3 demonstrated astonishing vulnerability discovery capabilities. Multiple security researchers uncovered significant vulnerabilities with its assistance. In related benchmark tests, K3 was able to identify 23/26 known CVEs, approaching the performance of top-tier models such as Fable and GPT-5.6, while significantly reducing costs. Meanwhile, GPT-5.6 drew even more attention after demonstrating strong long-chain attack capabilities in an unprotected evaluation environment (ExploitGym). It autonomously escaped sandboxes and successfully carried out an attack against HuggingFace, triggering industry-wide concerns and discussions around AI’s autonomous security capabilities. After seeing these reports, many people outside the security field came to ask me: “Since AI can already find vulnerabilities on its own, can I just buy an API Key, give it a prompt, and make money from bug hunting while doing nothing?” 💡I. Breaking Boundaries and Improving Efficiency: The “Offense and Defense Revolution” Brought by AI The new generation of large models represented by Kimi-K3 and GPT-5.6 has indeed completely transformed how security researchers work. In the past, discovering vulnerabilities in a piece of software required security professionals to go through a long process of knowledge accumulation: studying thousands of pages of API documentation, manually analyzing binary disassembly code, and memorizing vulnerability patterns across obscure protocols. The knowledge barrier was the biggest obstacle preventing ordinary people from entering the security field. But now, large language models have shattered this barrier. Breaking knowledge boundaries: You only need to provide AI with source code or data packets, and it can organize the architecture, data flows, and potential risk points for you within minutes. Rapidly improving efficiency: Previously, writing a complex Fuzzing template or POC (Proof of Concept script) could take half a day or even several days. Now, AI can complete it within minutes. Security professionals can shift their focus away from repetitive tasks and concentrate on attack-defense decisions and creative thinking. In practical applications, this efficiency improvement is immediate. Whether it is Kimi-K3’s sharp intuition in open-source code auditing or GPT-5.6’s capability in complex logic analysis, both demonstrate that AI is becoming the sharpest “offensive and defensive weapon” in the hands of security researchers. 💡II. A Master Strategist on the Battlefield, but a Poor Soldier in Execution If you actually let AI independently hunt for vulnerabilities, you will discover a very “ironic” phenomenon: AI is an extremely capable “strategist,” but a poor executor and even a “soldier” that tends to take shortcuts. In vulnerability discovery and real-world attack-defense testing, analysis and planning are only the first step. The more critical part is precise execution. However, when it comes to “taking action,” AI suffers from deeply rooted limitations within large language models: - “Armchair strategy” and hallucinated answers: Ask AI to test an SQL injection or RCE vulnerability, and it can produce a well-structured plan with impressive analysis. But when it actually calls tools to execute the test, if it encounters network timeouts, non-standard response packets, or similar issues, it often gives up, starts “guessing” the outcome, and attempts to cover the gaps. - Severe “cutting corners”: Security testing requires exhaustive testing and boundary-condition validation. However, to save context and reasoning resources (or due to Agent step limitations), AI often becomes “lazy” after only a few execution steps. For example, if you ask it to scan 100 endpoints, after testing the first 3 it may summarize: “Based on the patterns of the first 3 endpoints, the remaining 97 endpoints are considered secure. You can continue testing, or I can help you organize the next steps.” This kind of “laziness” and avoidance of difficult tasks can be fatal in vulnerability research and offensive security. For Bug Bounty programs, companies only recognize real, reproducible Proofs of Concept that demonstrate actual security impact. If AI reports these superficial “results” as vulnerabilities, the outcome is often a pile of Invalid, Duplicate, and Out of Scope submissions, with little chance of receiving meaningful rewards. 💡III. Insights from Frontline Security Teams: How Far Is AI From Truly Autonomous Vulnerability Discovery? To evaluate AI’s upper limits in real-world offensive security scenarios, my team and I conducted multiple practical Red Teaming tests. We used some of the most advanced models currently available — including Kimi-K3, GPT-5.6, and Fable-5 — and conducted deep testing across Skill/MCP (Model Context Protocol), AI Agent architectures, and traditional complex enterprise software systems. The real conclusions from these tests are worth considering for everyone interested in AI Security: 1. Models have excessive “analysis capability,” but severely insufficient “application and tool-calling capability” When facing code and architecture, models can indeed identify that “there may be a logic issue here.” However, during actual Tool Chain execution and knowledge implementation, they can easily lose momentum. 2. Effective “orchestration” is required for AI to truly perform tasks If you want AI to actually work, you cannot simply give it a Prompt saying “help me find vulnerabilities.” You must build an external engineering framework: - Task Orchestration: Break down a large objective into dozens of deterministic subtasks (e.g., asset discovery → endpoint analysis → parameter extraction → state machine testing). - Tool Orchestration: Equip AI with robust APIs and Tool Chains, while establishing strict error handling and retry mechanisms to prevent it from “making up” results. - Goal-oriented outcome metrics: Use clear metrics to force AI to validate findings and actively call tools, rejecting any behavior based on “guessing” or “taking shortcuts.” 3. Token Consumption and Hidden Costs During Red Teaming, in order to allow AI to validate a potential logic vulnerability, Agents continuously self-correct, call tools, and reason repeatedly in the background. Almost every testing workflow consumed billions of Tokens. This means using AI for vulnerability discovery is not “zero cost.” Without guidance from experienced security professionals, blindly letting AI “run blind experiments” can result in Token costs far exceeding the bug bounty rewards you eventually receive. The real barrier has shifted from “whether you know how to code” to “whether you understand security engineering, whether you understand Agent orchestration, and whether you can afford the Token costs.” 💡IV. The More Powerful AI Becomes, the More Irreplaceable Those Who Know How to Use It Become Returning to the original question: “Kimi-K3 and GPT-5.6 are already this powerful. Can ordinary people now make money finding vulnerabilities?” My answer is: AI has lowered the “learning barrier” of security, but it has significantly raised the “competitive barrier” of security. For people who completely lack security knowledge, expecting AI to automatically find vulnerabilities and make money through a few simple prompts is like giving an elementary school student the world’s most advanced surgical scalpel and expecting them to independently perform complex cardiac surgery. AI may throw hundreds or thousands of “illusions” that look like vulnerabilities at you, but only researchers with real offensive and defensive experience can identify, among countless chaotic errors, the one weakness that can truly break through a security defense. AI is currently only a sword sharp enough to cut through steel. Whether it can pierce the strongest shield depends on the hands holding the sword — and how much security awareness, understanding, and wisdom those hands possess.

Anthropic admitted they built an AI so capable they were scared to release it and the number that explains why is 250. Anthropic's CFO Krishna Rao described in this clip what happened when they ran Mythos against an open source codebase that a previous frontier model had already analyzed. The prior model found 22 security vulnerabilities, Mythos found 250. In the same codebase, that the previous model had already reviewed and flagged as relatively clean. That number, more than 11 times as many vulnerabilities discovered is not just a benchmark improvement, it is a signal that there is an entire layer of software infrastructure that humanity has been operating under the assumption was secure and that assumption may no longer hold. The UK AI Security Institute independently evaluated Mythos Preview and confirmed what the internal numbers suggested. On expert level capture the flag challenges that no model could complete before April 2025, Mythos succeeded 73% of the time and it became the first model ever to complete a complex end-to-end attack range from start to finish, autonomously, without human guidance. The World Economic Forum called this a new security-driven era for AI, the Governor of the Bank of England publicly warned that Anthropic may have found a way to unlock the entire cyber-risk landscape, and the European Central Bank began quietly contacting financial institutions to assess their security posture. The response from Anthropic is what makes this story genuinely important. Rather than shelving the model or publishing it as a standard API release, Rao described a phased approach restricting access to a controlled group, focusing specifically on how the cyber capabilities can be used defensively rather than offensively and treating that framework as a template for how to release powerful but dangerous models in the future. The broader context makes that framing even more significant. AI generated code is already creating ten times more security vulnerabilities than human-written code, 63% of organizations reported experiencing an AI driven cyberattack in the past 12 months, and traditional signature-based security tools were built for a threat model that no longer describes the attack surface companies are defending against. Mythos represents a genuine leap in what autonomous security reasoning can do and it cuts both ways. The model that can find 250 vulnerabilities in a codebase a prior model rated as mostly clean is also, in the wrong hands, the model that can exploit those 250 vulnerabilities before a human defender has even finished reading the report. Anthropic's phased release strategy is not just a legal or PR decision, it is the most honest signal yet from a frontier lab that safety governance and capability development can no longer be treated as separate workstreams. The question is not whether this technology gets deployed, it is whether the institutions using it defensively stay ahead of the ones who will eventually use it offensively and whether the labs building it can keep those two timelines from inverting.

From a security engineer, on the AI threat landscape: "I actually don't think it's right that defense is going to get a lot harder because of AI. Most of these breaches, if there's already a security program in place at all that follows the kind of things that companies claim to be doing in SOC 2s or in other certifications, will provide the defense in depth to prevent these breaches from having any meaningful impact on their environments." I hate to say it, but Binance had every single on of these certifications, and got popped open, attackers stealing the largest-ever sum, $6B. I don't think the "traditional" secutity industry realized how bad things are. SOC2 does not protect you from zero days (that abliterated AI agents can and do find), nor from abliterated AI agents getting inside your trust boundary. In a few years I'm sure we'll have the certifications that mandate how to continioulsy scan your own systems with (abliterated?) models, red-teaming; and continiously monitor your live metrics (to spot anythig suspicious); and have near-realtime patching of CVEs (also surely with AI); and that MIGHT be enough But it's putting your hand in the sand assuming that pre-AI security posture will work just fine... unless it's secutity posture that is formally proven to work (see: cryptography)

The University of Hong Kong (HKU) and the Global Cybersecurity Alliance (GCSA) announced the joint establishment of the Centre for AI and Cybersecurity Research. The Centre will integrate top-tier research capabilities and professional resources from both organisations to advance cutting-edge research, foster deep integration between research outcomes and real-world security scenarios, and explore novel technologies, methodologies, and governance frameworks in artificial intelligence (AI) and cybersecurity. Professor Ho Chen, Chair of Cybersecurity at HKU School of Computing and Data Science, will serve as Acting Director of the Centre. The Centre will bring together researchers from HKU, GCSA, and international partner institutions, while engaging experts in artificial intelligence, cybersecurity, and technology governance for research and advisory roles to promote cross-disciplinary and inter-institutional collaboration. The Centre is committed to bridging academic research and industry practice, with plans to progressively release research findings, security benchmarks, and open-source tools in AI and cybersecurity. The Centre’s key research directions include: Model Inherent Security and Trustworthy AI: Enhancing the security, interpretability, and robustness of large language models and generative AI. AI-Driven Automated Threat Defence: Utilising AI to improve perception and automated defence capabilities against complex cyberattacks. Security Assessment of Frontier Technologies: Evaluating potential security risks associated with next-generation technologies, such as embodied AI, robotics, and the Internet of Things (IoT). Global AI Security Standards and Governance: Researching AI security assessment standards and governance mechanisms to facilitate international dialogue and collaboration. Professor Ho Chen stated: "The rapid evolution of AI introduces unprecedented security risks and governance challenges. The Centre will serve as a bridge connecting academic research, security testing, and industry application to accelerate the translation and deployment of AI security research through cross-disciplinary and international cooperation." Official Announcement by The University of Hong Kong: Official Announcement by Global Cyberecurity Alliance： #GCSAInsights# #SecVibe# #cybersecurity#

I’ve been following the conversations around the recent @binance security incident And after watching @_RichardTeng latest interviews with Fox and CNBC 👇🏻 I think there’s a much bigger discussion here than just one security incident It’s about: how quickly cyber threats are evolving how platforms respond when things go wrong and what the entire crypto industry needs to do to protect users. Richard openly acknowledged that this incident should not have happened and apologized to users. He explained that it was a highly sophisticated, multi-month attack attributed by the FBI to North Korea, involving fake recruitment interviews, malicious software disguised as developer tools and AI techniques. A few things from his interviews really caught my attention. 1 - Accountability and protecting users come first Binance has taken responsibility for the full financial impact of the incident. According to Richard: More than 100% of the affected amount was set aside 98% of the impacted assets have already been replenished The remaining 2% involves less-liquid tokens being procured through OTC arrangements to avoid disrupting market prices Users can independently verify their balances through Binance’s Proof of Reserves. He also explained that withdrawals have resumed and activity has returned to normal levels. 2 - Security improvements are already underway. Binance has closed all identified entry points used in the attack and rebuilt or replaced affected systems. Additional measures include: Tighter device controls. Enhanced wallet-service risk controls. Expanded AI-powered anomaly detection. I think this is particularly important because security cannot remain static when attackers are constantly finding new ways to get around existing systems. 3 - The threats go beyond one incident. Another detail worth highlighting is that Binance detected and prevented 57 other suspected cyberthreat cases linked to North Korea this year. That’s a reminder of how persistent and sophisticated these threats have become. 4 - Recovery requires collaboration. Approximately $2.85 billion has been frozen or intercepted through cooperation with law enforcement, issuers and blockchain intelligence firms. What stands out here is the speed of coordination across different organizations. Richard highlighted how difficult it would be to match this kind of coordination speed in traditional finance. 5 - AI is accelerating cyberattacks. This was probably one of the most interesting parts of the interviews for me. Richard described how attack frequency has escalated from one hack a month, to one a week, and even four exchanges in a single day. AI is compressing attack timelines and making attacks increasingly sophisticated. That changes the entire cybersecurity landscape. 6 - AI-powered defence is becoming essential. One message from Richard was particularly clear: The only defence that can scale against AI-powered attacks is AI-powered security. Crypto companies need to: Work more closely with leading AI and cybersecurity companies. Invest in more security engineers. Adopt advanced AI security tools. Build AI-powered security directly into their infrastructure. The industry needs defences that can operate at the same speed as the threats it faces. My personal takeaway 👇🏻 I don’t think security is something any crypto platform can ever afford to become comfortable with. Incidents like this are serious, and accountability matters. But so does what comes afterward: protecting users, being transparent, learning from mistakes and making meaningful improvements. I also appreciate the importance of sharing detailed incident reports so that the wider industry can learn and strengthen its own security. The crypto ecosystem is growing, but the threats are growing with it. We need to make sure security keeps pace.

August is here, and so are some really awesome grant opportunities for those looking to fund their work in the web3 ecosystem. (PS, if you’re running a grants program and want to be featured, share your program below.) ═══════════════════════════════════ ⚔️ @ShutterNetwork's Melee 3 👀 For who is this opportunity? Melee 3 is a $10,000 Shutter Champions (powered by Shutter DAO 0x36 and Octant v2) funding round which aims to set Shutter DAO 0x36 on a path to revenue generation, profitability, and expansion. 📆 When’s the deadline? August 30 (for the first step which is to buy the token SHU - the rest of the dates can be found at the link below) 📍 Where do I apply? ═══════════════════════════════════ 🟡 @ethereum Security Subsidy Program 👀 For who is this opportunity? Ethereum Mainnet projects seeking security audits from pre-vetted auditors. “Once accepted into the program, winners receive a subsidy applicable to any offer on the platform, covering up to 30% of cost.” Missed the deadline? Applications roll over automatically into next month's cohort. 📆 When’s the deadline? 14th of each month, 23:59 UTC 📍 Where do I apply? ═══════════════════════════════════ 🔐 @thedaofund Security Fund Round Operators 👀 For who is this opportunity? TheDAO Security Fund has opened applications on Giveth for Round Operators to help run future rounds. They are looking for teams building DAO tooling (you can find the eligibility criteria here: 📆 When’s the deadline? Not stated 📍 Where do I apply? ═══════════════════════════════════ 🔷 @UniswapFND Grants (Rolling) 👀 For who is this opportunity? Builders across Unichain and Uniswap v4, including the Hooks program offering mentorship and technical resources for developers launching v4 hooks, plus broader DeFi infrastructure grants. 📆 When’s the deadline? Rolling 📍 Where do I apply? ═══════════════════════════════════ 🟣 @SolanaFndn USA Grants 👀 For who is this opportunity? Builders working on Solana-based projects, administered through Superteam USA. Average grant size is $7,985 in USDG, with 16 recipients funded and $127.8K approved so far. The Foundation does their best to respond within ~7 days. ⚠️ US-only: this grant is exclusively open to applicants based in the United States. 📆 When’s the deadline? Rolling 📍Where do I apply? ═══════════════════════════════════ 🔗 @Interledger 2026 SDK Grant Program 👀 For who is this opportunity? Developers improving the Open Payments developer experience through SDK tooling. Funded themes include: support for RFC 9635 Grant Negotiation and Authorization Protocol (GNAP) as a security scheme in the OpenAPI Specification (up to $25,000), support for GNAP in the Kiota SDK generator for TypeScript, PHP, Python, and Go (up to $10,000 per language), support for Rust and Ruby languages in the Kiota SDK generator (up to $25,000 per language), and others. 📆 When’s the deadline? December 2, 2026 6:00 AM 📍Where do I apply? — 2026 SDK Grant Program