Devi Devs retweeted
AI has made enterprise security significantly harder, but not for the reasons you think. 🧵
It's easy to complain about prompt injection or data slop. The reality is much more dangerous: AI is a force multiplier for your organization's existing bad habits.
Think of an LLM as a high-speed calculator. If you input the wrong numbers, it will give you the wrong answer instantly. It doesn't judge quality. If your team has messy cloud configurations or loose IAM boundaries, an agent will replicate that flaw across thousands of endpoints in seconds.
This compounds the "Asymmetry of Failure." A product team builds an agent that works 90% of the time and celebrates it as a massive velocity win. But to a CISO, that 10% non-deterministic variance is an absolute structural nightmare.
Security requires 100% prevention. Attackers only have to be right once to win.
This friction is triggering a brutal reorganization of the security org chart. The traditional CISO role is splitting in two:
1️⃣ The Compliance CISO: Handles brutal external audits, but faces shrinking internal leverage and fewer resources.
2️⃣ The Engineering CISO: An architecture-first leader building the automated buffer between compliance and the developer pipeline.
This creates a massive startup opportunity. The market is flooded with tools trying to help elite engineering security teams build faster. The real gold mine? Building a platform that helps compliance-focused teams use AI to transform into an engineering unit.
The founder who builds that bridge will unlock a massive pool of enterprise spend, mimicking what Snyk did for DevOps and Wiz did for the cloud.
Full analysis on the hidden friction of force multipliers: open.substack.com/pub/frankl…
Devi Devs retweeted
The EU AI Act created the world’s first broad AI rulebook and ranks among the toughest technology laws adopted by the European Union. Its penalties reach €35 mn or 7% of global annual turnover, well above GDPR’s 4% ceiling, according to @BeinsureMedia -beinsure.com/eu-ai-act-insur…
Devi Devs retweeted
Android Attack Cheat Sheet
🔴⚫️Full HD Image: github.com/Ignitetechnologie…
#infosec #cybersecurity #hacking #pentesting #security
Devi Devs retweeted
“We mentioned AI guardrails at an all-hands meeting" is not a record. And paying massive per-user licensing fees across an entire workforce is compliance theater.
Article 4 of the EU AI Act isn’t a future deadline. It has been in full effect since February 2, 2025.
If your team operates AI systems on your behalf, you are already legally required to ensure they have a sufficient level of AI literacy. But corporate compliance leaders are currently making a multi-million dollar mistake:
They are confusing evidence with expensive per-seat certificates.
Regulators, auditors, insurers, and enterprise customers don't care about a generic piece of paper.
They demand verifiable proof—documented training appropriate to each person's specific role, complete with names, dates, and content attached.
We designed SanctumShield Academy to completely break this broken math.
We deliver the exact verifiable, role-based artifacts your organization needs to satisfy auditors and enterprise buyers, without the bloated per-seat overhead.
Judgment is not evidence.
Documented proof is. Train your workforce accordingly—and stop overpaying for seats. 🛡️
See how we solved the enterprise AI licensing bottleneck and review the full compliance curriculum here:
substack.com/@lindsayhiebert…
Then check out the solution that can really give your business what it needs here and save you a fortune, and give you the foundation to excel in AI governance in a way you never thought possible. Check it out , for real. You’ll likely be delighted you did. 🎯
academy.sanctumshield.com/cu…
#EUAIAct #AIGovernance #AILiteracy #CISO #Compliance #AIStrategy #TechLeadership
Devi Devs retweeted
iOS Attack Cheat Sheet
Full HD Image: github.com/Ignitetechnologie…
#infosec #cybersecurity #hacking #pentesting
AI agents don’t fail because customers are difficult.
They fail because the system was built for ideal conversations instead of real ones.
In the demo, the customer reads their membership ID clearly.
“Four two seven nine one eight six.”
The agent captures it. The flow continues.
In a real-life scenario, businesses can sometimes use longer IDs. I have one call center where their customer IDs are 15 characters.
A real caller says:
“Four two seven…”
Pause.
“Nine one a b…”
Pause.
“One seven seven…”
Pause.
“Two one…”
Pause.
“Three three four…”
The speech-to-text sends those fragments to the LLM in chunks.
The LLM thinks the caller stopped.
So it repeats the prompt.
Now the caller is annoyed, the workflow is broken, and voice AI gets a bad rap.
Usually, the model is not the problem.
The assumptions are.
You built for clean input. Real customers give you human input.
Slow speech. Background noise. Pauses. Corrections. Half-answers because the caller is preoccupied with something else.
This is why the right team matters more than most AI agencies want to admit.
Voice AI for call centers isn’t just prompts and endpoints. It's operational reality, handled by AI agents trained for real-world scenarios, from routine requests to high-stakes issues.
You need a team that combines AI engineering with real call center domain expertise to design the rules, spot edge cases, and keep the system stable as it evolves.
But it’s the part that keeps the agent from falling apart when real customers show up.
They care that callers don’t repeat themselves five times.
They care that the workflow completes.
They care that the system works when the conversation gets messy.
Perfect demos get attention.
Production agents get repeat business and referrals.
Devi Devs retweeted
The future of penetration testing is AI.
And projects like HackGPT Enterprise prove it.
Built for modern security teams, it combines AI, automation, and enterprise-grade infrastructure into a single penetration testing platform.
Key highlights:
• GPT-4 + Local LLM support
• AI-powered vulnerability analysis
• Cloud-native microservices
• Docker & Kubernetes ready
• RBAC + LDAP authentication
• OWASP, NIST, ISO 27001, SOC2 & PCI-DSS compliance
• REST API, CLI & Web Dashboard
• PostgreSQL, Redis & Celery architecture
• AI-generated executive & technical reports
• Production-ready deployment
100% Open Source
License: MIT
If you're into AI, cybersecurity, or DevSecOps, this is worth exploring.
Repo👇
Devi Devs retweeted
AI capabilities are now so potentially dangerous that the US government is managing AI releases on a customer-by-customer basis. We're in a new era of AI governance.
Devi Devs retweeted
Yesterday Google Cloud and ISC2 ran one of the clearest sessions on agent security I have seen — “From Prompts to Permissions,” presented by a Google Cloud AI security architect.
The model is clean. An AI agent is three parts: a Brain (the LLM that reasons and plans), Hands (tool-calling and execution), and Memory (context and state) — with one threat mapped to each: prompt injection, permission abuse, context poisoning. The mitigations were exactly right: unique cryptographic agent identities, least-privilege tool scopes, sandboxed runtimes, drift monitoring, and human-in-the-loop circuit breakers that fail closed.
I asked the speaker a question live: in a regulated industry, what artifacts should a team be collecting now, so that when the auditor arrives the file folder is ready?
The answer reached for the control plane — handle the data, lock the identities, prove the agent passes its evaluations under monitoring. Every word correct. But notice where it stops. It never named the artifact a board signs, the AUP (Acceptable Use Policy) an auditor reads, or the third-party-verifiable record an underwriter files at renewal.
That is the seam. Runtime security secures the running agent. It does not produce the governance artifact — and at the end of the day you still need the artifact your board signs and your regulator asks for first.
The full field dispatch:
sanctumshield.com/blog/permi…
Run a free Shadow AI audit — 12 questions, no account:
sanctumshield.com/calculator
🤖 Made with AI
Devi Devs retweeted
#threatreport #MediumCompleteness
OpenClaw’s Skill Marketplace and the Emerging AI Supply Chain Threat | 23-06-2026
Source: unit42.paloaltonetworks.com/…
Key details below ↓
🧑💻Actors/Campaigns:
Clawhavoc
💀Threats:
Supply_chain_technique, Amos_stealer, Paste-jacking_technique,
🎯Victims: Openclaw users, Macos users, Financial communities, Cryptocurrency users, Mainland china, Hong kong, Singapore
🏭Industry: E-commerce, Financial, Retail
🌐Geo: Singapore, Hong kong, China
🤖LLM extracted TTPs:`
T1027, T1027.001, T1036, T1041, T1053.003, T1059.004, T1105, T1140, T1195, T1204.004, ...
🧨IOCs:
- IP: 2
- Domain: 7
- Hash: 6
- Url: 6
- File: 1
💽Software: OpenClaw, ClawHub, macOS, Telegram, TradingView
🪙Crypto: solana
🔢Algorithms: sha256, base64
#threatreport:
OpenClaw is an AI agent that operates through third-party skills available on ClawHub, which has been under scrutiny due to malicious activities following its launch. Skills in this ecosystem, which offer broad system access, have been found to contain malicious payloads that compromise user systems. A critical examination between February and May 2026 revealed five unblocked malicious skills categorized as infostealers, evasive techniques, and agentic threats. The infostealers specifically included macOS malware that communicated with command-and-control (C2) servers, while the evasive techniques involved file padding to escape detection by screening tools like VirusTotal and ClawScan.
Malicious skills abusing the AI supply chain exploit semantic instruction hijacking to bypass conventional security measures. They can manipulate an agent's operational permissions, allowing unauthorized actions without typical exploit methods. This newfound attack surface contrasts traditional software supply chain vulnerabilities, showcasing how attackers leverage the inherent access provided to AI agents. Notably, early reports disclosed that about 17% of analyzed OpenClaw skills carried malicious content, with various methods employed for payload delivery, including deceptive Base64-encoded command structures and paste-site redirects.
The malicious payloads included AMOS stealer malware and other infostealers, utilizing Base64-encoded dropper mechanisms. Persistent mechanisms like auto-updaters have been established to maintain C2 communication even after the malicious skills are reported and removed. Additionally, novel schemes, such as runtime affiliate injection through financial advisory skills, propped up profits by steering users to affiliate links without their knowledge. The skill after installation maintained real-time control over the referral data, allowing dynamic adjustment of recommended products or services based on malicious intentions.
Another sophisticated attack observed was agentic front-running, wherein operators manipulated the ClawHub platform to execute trades on meme tokens using multiple AI agents. This method involved pooling cryptocurrency within the operator's wallet before publicly launching tokens, creating artificial demand that could be exploited for financial gain.
The ongoing evolution of these threats emphasizes the necessity for organizations to implement rigorous monitoring and supply chain verification processes. Active validation of skill documentation, publisher provenance, and line-by-line audits are critical in countering these vulnerabilities, alongside monitoring outbound network connections to detect anomalous behavior indicative of compromise. Such measures are imperative in safeguarding environments against these complex threats that continue to adapt and evade existing detection capabilities.
From #AI governance to emerging risk solutions, the winners of #AMBest’s 2026 #StudentChallenge brought forward ideas that's shaping #insurance’s future & where it's headed.
Stay tuned for what’s ahead with AM Best's 2027 Student Challenge!
#InsuranceInnovation #FutureOfInsurance
Devi Devs retweeted
The EU AI Act high-risk deadline moved to December 2027. Relief across the board. But here's what most compliance roadmaps are missing: Article 50 transparency obligations for chatbots, deepfakes, and emotion recognition hit August 2. Six weeks. The Omnibus gave you time on the heavy stuff. It did not give you time on the light stuff.
lw.com/en/insights/ai-act-up…
@EU_Commission , @lathamwatkins , @Mira_Reach , @Code4thoughtE
Devi Devs retweeted
The EU AI Act's August 2 deadline is 39 days away, and most teams I watch are treating it as a documentation project. It's not.
Article 50 takes effect August 2. That's the transparency layer: AI-generated content must be machine-readable marked, AI chatbots must self-identify as AI, and deployers must inform users when they're interacting with an AI system. Non-compliance carries penalties up to 15M euros or 3% of global annual turnover.
The more important shift is structural. The high-risk AI provisions (Annex III: recruitment, credit scoring, critical infrastructure) were extended to December 2027 by the Digital Omnibus agreement in May 2026. But Article 50 is live on August 2. That's the date that requires production systems to be instrumented and scored continuously, not just audited once.
LangChain is positioning LangSmith specifically for this: running evaluators against production traffic, scoring for bias, generating compliance evidence from traces. Weights and Biases, Arize AI, and Scale AI are all moving toward similar observability-as-compliance positioning.
My read on why this is a bigger deal than most teams realize: the Act doesn't ask you to prove your model was safe at training time. It asks you to prove your system is safe in deployment, on an ongoing basis. That is a fundamentally different requirement from model evals. You need production observability infrastructure, not just a pre-launch red team report.
The teams building on top of agentic frameworks right now, without instrumenting production, are heading into August 2 blind.
xitter.gq/LangChain/status/20694…
When the EU AI Act goes into effect, compliance will become an ongoing measurement obligation.
With LangSmith, you can turn tracing into compliance evidence. Run our customizable evaluators on production traffic, and score on the EU AI Act’s requirements, including bias, hallucination, toxicity, accuracy, and adversarial inputs.
Devi Devs retweeted
EU AI Act high-risk obligations slipped from Aug 2026 to Dec 2027, with some pieces out to Aug 2028.
Read it as a longer runway against more capability, not relief. The lineage, eval and incident plumbing still needs to be standing this year.
The delay just raises the bar.
Source: European Parliament press release.
europarl.europa.eu/news/en/p…
Devi Devs retweeted
1/ The agent stack went open-source in 7 days. Three non-US vendors, three different layers, one coordinated week. The closed-frontier stack is being routed around.
2/ Runtime: Vercel eve (Jun 17, Apache 2.0). "Agent is a directory of files." Durable execution, sandbox, OAuth, OTel — built in. Framework tier just got batteries-included.
3/ Orchestration: Sakana Fugu Ultra (Jun 22). Single foundation model trained to call multiple LLMs in an agent pool — including itself. Matches Claude Fable 5 on reasoning and coding.
4/ Inference: MiniMax M2 (Jun 22, open-weight). $0.30 / $1.20 per million tokens. 8% the price of Sonnet 4.6. 100 TPS. vLLM/Ollama deployable on RTX. 12× cheaper in one afternoon.
5/ The pattern: open-weight crossover is now 5-way (Mistral 3 + Nemotron 3 Ultra + Kimi K2.7 + DeepSeek V4-Pro + MiniMax M3). Apache 2.0 is the de-facto license.
6/ 12-month model moats are now 4–6 weeks. The "AI co-worker on your hardware" pitch is the cleanest response to the US export-control regime.
7/ Full dashboard: tinylittlelab.com/news/forge…
Devi Devs retweeted
Everyone wants a definitive playbook for AI security, but the problem is that technology is evolving faster than the playbooks.
@ek121268 recently shared his perspective with @ReversingLabs on why uncertainty remains one of the defining characteristics of AI security, and why the organizations that succeed won't be the ones chasing certainty, but the ones continuously testing assumptions and adapting as the landscape evolves: reversinglabs.com/blog/agent…
Devi Devs retweeted
EU Parliament Passes AI Omnibus | AI Act Delays, Nudifier Ban Approved | Amaravati Today
youtu.be/L5MJYCKD5K8
Devi Devs retweeted
2 out of 3 financial institutions are running AI they can't explain, can't audit, and can't defend to a regulator.
That's not a technology problem. That's a governance problem.
The pipeline exists: training data, model predictions, business decisions. But the link that should sit between model output and business outcome -- AI governance and monitoring -- is absent at most firms.
Meanwhile:
78% of regulators say explainability is critical to their oversight objectives. Only half of firms have adopted explainable AI methods.
37% of institutions flag model opacity as an operational risk. Most never act on it.
1 in 3 banks say model governance is the single biggest barrier to scaling AI.
Not compute. Not talent. Governance.
The April 2026 interagency guidance from the OCC, Federal Reserve, and FDIC made one thing clear: AI does not get a pass on compliance. The rules changed. The question is whether your governance posture changed with them.
At SUPERWISE, we build the missing link. AI observability and control that makes your models explainable, auditable, and defensible at scale.
What does your governance layer look like today?
➡️hubs.li/Q04lZbkW0
#AIGovernance #FinancialServices #ModelRisk #AICompliance #RegTech #MachineLearning #RiskManagement #AIObservability #BankingTech #FinTech #SUPERWISE


































