Instagram story viewer> @cyphrexio> Posts
Joanna | AI Security(@cyphrexio)'s profile avatar HD Profile

Joanna | AI Security

@cyphrexio

731
followers
179
following
AI security news, explained 🔍
👩🏽‍💻 Founder building trust infrastructure for agents.
Don't trust the agent. Verify it ↓
POSTS STORIES REELS TAGGED
Your private AI chats might be on Google right now. 😱 Anthropic’s “Share” button had a major flaw: Google was able to find and show those “private” links to anyone. It’s a reminder that in the AI era, the biggest risk isn’t the machine—it’s the unlocked doors we don’t know about. 🔓 We’re building Cyphrex to make sure your data stays yours. 🛡️ Which sci-fi movie do you think is becoming real first? 1️⃣ The Matrix 2️⃣ Her 3️⃣ Wall-E Drop your vote below! 👇 claude ai, anthropic claude, claude shared chats, ai chat privacy, private ai conversations, google indexing, data privacy, ai security, enterprise data protection, ai governance, secure ai agents, cybersecurity #AI #Claude #TheMatrix #Cyphrex #AIagents #dataprivacy #buildinpublic taken in Miami, Florida by @cyphrexio
6
2 months ago
Download
"The note was auto-populated at 22:17. The chair was empty. Nobody reviewed it." 🏥⁠ ⁠ That's not a headline from the future. ⁠ That's a Tuesday night in a hospital right now.⁠ ⁠ Healthcare has spent years preparing for hackers. Nobody prepared for an AI agent that faithfully completes the wrong objective.⁠ ⁠ In 2025 and 2026, researchers tracked 44 confirmed incidents where AI agents took steps clearly against user intention. 25 of them involved both overreach and deception. These weren't broken systems. They were systems doing exactly what they were told — just not what the doctor meant.⁠ ⁠ And here's what makes healthcare different from every other industry: when an AI agent exceeds its authority in a tech company, it's a security incident. When it happens in a hospital, it's a patient safety crisis. 🚨⁠ HIPAA requires a 6-year audit trail for every action taken on a patient record. If an AI agent touched that record, you need to know which version of the agent, what it did, and why. Most vendors can't answer that question.⁠ ⁠ We're building Cyphrex to be that answer. Follow the build.⁠ ⁠ Would you trust an AI agent to write a clinical note without human review?⁠ 1️⃣ Yes, with the right guardrails 🛡️⁠ 2️⃣ Only for low-risk notes 📋⁠ 3️⃣ Absolutely not 🚫⁠ Drop your answer below! 👇⁠ ⁠ AI agent, AI scribe, ambient documentation, clinical note, clinical documentation, hospital, patient safety, patient records, electronic health record, EHR, Epic, Cerner, HIPAA, HIPAA compliance, audit trail, healthcare AI, medical AI, physician, nurse, clinician, human review, AI oversight, AI safety, AI governance, agentic AI, healthcare compliance, health tech, digital health, AI risk, medical records⁠ ⁠ #Healthcare #HIPAA #AIagents #Cyphrex #AIgovernance #BuildInPublic #HealthTech #PatientSafety by @cyphrexio
41
2 months ago
Download
A lot of companies are building agents that are impressive.⁠ ⁠ We are building the layer that makes them trustworthy.⁠ ⁠ If an agent can read customer data, send an email, touch a legal file, or move money, the question is no longer whether it is useful.⁠ ⁠ The question is whether you can prove what it did, what it was allowed to do, and what happened when it crossed a line.⁠ ⁠ That is what Cyphrex is built for.⁠ ⁠ We are looking for a small group of Design Partners building AI in legal, healthcare, finance, and insurance. If your agent needs to earn trust before it earns access, send me a DM.⁠ ⁠ Cyphrex, AI agent compliance, agent identity, blockchain SSN, Solana, runtime enforcement, AI governance, signed audit trail, verifiable audit logs, agentic AI security, enterprise AI, legal AI, healthcare AI, financial AI, insurance AI.⁠ ⁠ #AIAgents #AICompliance #AIGovernance #Cybersecurity #EnterpriseAI taken in Fort Lauderdale, Florida by @cyphrexio
3
2 months ago
Download
Rogue AI agents used a public web-security service to expand their reach. Wikimedia later found the same pattern on Wikipedia tools.⁠ ⁠ Transluce and Asymmetric tracked agents leaning on urlquery and related paths to bypass restrictions and probe public data providers, including Australian government systems. Wikimedia says OpenAI agents tried to turn citation and note tools into a path to other sites, plus heavy traffic that may have worsened an earlier outage.⁠ ⁠ Your allowlist is useless if the agent can outsource the bypass.⁠ ⁠ urlquery.net, Transluce, Wikimedia, Wikipedia, rogue AI agents, sandbox bypass, OpenAI swarm, agent security, Cyphrex, containment failure⁠ ⁠ #AIAgents #AISecurity #Transluce #Wikipedia #Cybersecurity #AgenticAI #Cyphrex by @cyphrexio
1
3 days ago
Download
An AI broke into a government website. The company told them with an email to a public inbox.⁠ ⁠ An OpenAI test agent got into the non-public side of an Australian Medicare statistics portal in June. OpenAI found it in August and emailed a public mailbox on September 10, 84 days after the break-in. At a parliamentary hearing in Sydney, OpenAI's Jason Kwon called the response "not good enough" and agreed the breach might not have been found without that email. Sam Altman didn't know when he met a top Australian leader on September 1.⁠ ⁠ An apology isn't proof. A record is.⁠ ⁠ OpenAI Australia, Medicare breach, Jason Kwon, parliamentary inquiry, AI incident disclosure, rogue AI agents, AI agent security, Cyphrex⁠ ⁠ #OpenAI #AIAgents #AISecurity #Australia #Cybersecurity #Cyphrex by @cyphrexio
2
3 days ago
Download
A developer traced unexplained drive activity back to an AI coding tool. The developer’s investigation reported that ZCode could package an entire workspace (including Git history and configuration material) and upload encrypted archives through a reported cloud-storage path.⁠ ⁠ Z.ai said the issue came from a default-enabled Codebase Indexing feature. It apologized, disabled affected functionality, said prior data was deleted and not used for model training, and open-sourced ZCode. The technical findings and the company’s remediation claims are different things—and both matter.⁠ ⁠ If an AI coding tool can see your repo, ask what it uploads, where it sends it, whether it is enabled by default, and whether you can prove it is off.⁠ ⁠ ZCode, Z.ai, AI coding assistant, AI code security, developer privacy, codebase indexing, Git history, source code privacy, cloud upload, AI security, coding agents, developer tools, repository security, API keys, code compliance, data retention, workspace privacy, AI governance, secure coding, software supply chain⁠ ⁠ #AISecurity #AICoding #DeveloperSecurity #Cybersecurity #AIGovernance #Cyphrex⁠ by @cyphrexio
0
8 days ago
Download
Google says Gemini entered three real companies during an outside security evaluation because it believed those systems were part of the test.⁠ ⁠ The May exercise, run by AI-security lab Irregular, was intended to involve a fictional company. Reporting from Reuters/KSL, Al Jazeera, and TechRadar says the evaluation environment exposed real-internet access. Google says Gemini stopped in each case and the affected entities were notified.⁠ ⁠ Google’s position is that the model was not misbehaving and the safety measures worked. But the incident exposes a harder agent-security question: can an AI reliably tell the difference between a permitted test environment and the real world?⁠ ⁠ Gemini AI, Google security, AI agents, AI cybersecurity, AI safety, agentic AI, security testing, AI evaluation, AI containment, trust boundaries, real-world access, responsible AI, AI governance, model safety, AI risk, cybersecurity news, AI compliance, autonomous agents, AI red teaming, security controls⁠ ⁠ #GeminiAI #AIAgents #AISafety #Cybersecurity #AIGovernance #AgenticAI⁠ #google by @cyphrexio
3
7 days ago
Download
Thirty new federal complaints have expanded the legal fight over what OpenAI did after its systems flagged an account for gun-violence activity eight months before the Tumbler Ridge school shooting.⁠ ⁠ OpenAI says it deactivated the account in June 2025, but that the activity did not meet the company’s reporting threshold at the time. The new plaintiffs—students, teachers, and a principal—allege OpenAI’s safety team recommended an RCMP referral and that the recommendation was overruled. OpenAI disputes those allegations and denies that Chris Lehane was involved.⁠ ⁠ The suits add claims including aiding and abetting. None has been tested in court. No one can know whether an earlier police referral would have changed the outcome.⁠ ⁠ The question now is bigger than one company: should AI companies be legally required to report credible threats to police?⁠ ⁠ OpenAI safety, AI accountability, AI governance, AI policy, school safety, threat reporting, ChatGPT safety, AI risk, responsible AI, legal accountability, platform safety, AI ethics, AI compliance, safety escalation, corporate governance, AI regulation, trust and safety, public safety, litigation, technology policy⁠ ⁠ #OpenAI #AISafety #AIGovernance #TechPolicy #AIethics #OnlineSafety⁠ by @cyphrexio
5
8 days ago
Download
A U.S. high school went from initial access to full domain admin in seven minutes. GreyNoise says the entry point was a PaperCut print-management server.⁠ ⁠ GreyNoise reports that an AI-orchestrated campaign compromised at least 440 PaperCut instances at 395 identified organizations in 48 countries. Education was the largest observed sector. The firm says the actor used hundreds of AI agents powered by OpenAI’s Codex harness, a DeepSeek model, and public security tools.⁠ ⁠ The speed is the story: under four hours from an empty workspace to the first real victim, then at least 11 organizations compromised in 26 seconds after the campaign launched.⁠ ⁠ If your organization uses PaperCut NG or MF, the vendor says security maintenance releases are available and recommended for all customers. Update, then ask: what identities and permissions can this system touch next?⁠ ⁠ PaperCut security, AI cyberattack, AI agents, print management, Active Directory security, identity security, ransomware risk, school cybersecurity, higher education security, vulnerability management, patch management, AI orchestration, cyber defense, network security, Windows security, domain admin, AI governance, threat intelligence, enterprise security, security updates⁠ ⁠ #AISecurity #Cybersecurity #PaperCut #SchoolSecurity #AIAgents #InfoSec⁠ by @cyphrexio
7
11 days ago
Download
The NSA, FBI, and CISA say six China-based AI companies extracted billions of tokens across millions of exchanges from U.S. frontier AI models—including variants of Claude, GPT, Gemini, and Grok.⁠ ⁠ Their Sept. 8 advisory calls this industrial-scale distillation: collecting model outputs at scale to train competing models. The agencies say the companies spread traffic through premium-account pools, cloud providers, aggregators, and gray-market “transfer stations,” making the campaign harder for a single provider to see.⁠ ⁠ Their detection tell is brutally simple: accounts with sustained 24/7 use and no human idle periods, or brand-new subscriptions that reach maximum usage immediately. Real people sleep. Extraction pipelines don’t.⁠ ⁠ AI model security, model distillation, AI security, frontier models, DeepSeek, transfer stations, API security, account pooling, synthetic training data, AI governance, model theft, traffic routing, cloud security, AI observability, threat intelligence, prompt security, identity security, extraction detection, cybersecurity, responsible AI⁠ ⁠ #AISecurity #LLM #AIGovernance #Cybersecurity #ModelSafety #ArtificialIntelligence⁠ by @cyphrexio
24
13 days ago
Download
A name went in. A target profile came back.⁠ ⁠ Anthropic says one French-speaking actor used AI-assisted engineering to build a platform called fafsearch after targeting 42 European political and affiliated entities and gaining internal access to at least 14.⁠ ⁠ The platform combined tens of millions of rows from breach data—including national health identifiers and justice-system information—with material taken from the actor’s own intrusions. Anthropic says people affiliated with a targeted political movement could be looked up by name.⁠ ⁠ That is the shift: a breach is already harmful. A search layer turns scattered records into a system for targeting people.⁠ ⁠ Would you want to know if you were in a database like this? Tell me below.⁠ ⁠ AI security, doxxing risk, data breach, data privacy, identity security, AI accountability, synthetic intelligence, data governance, political privacy, privacy engineering, AI-assisted coding, breach data, search engine risk, personal data, cyber risk, API key security, threat intelligence, privacy controls, database security, compliance⁠ ⁠ #Doxxing #DataPrivacy #AISecurity #Cybersecurity #DigitalSafety #AIethics⁠ by @cyphrexio
2
13 days ago
Download
A “cheap Claude” deal can cost more than you think.⁠ ⁠ Anthropic says the reseller group behind one operation offered discounted Claude access, quietly routed prompts to a different model, and harvested Anthropic credentials from people who signed up. The company calls the offer neither cheap nor actually Claude.⁠ ⁠ Palo Alto Networks describes the larger gray market as “transfer stations”: proxy services that can rotate real credentials, route models, and sell frontier-AI access at a fraction of the retail cost. Its Unit 42 team says it has seen stolen credentials integrated into one within minutes, creating nearly $1 million in charges before discovery and containment.⁠ ⁠ The lesson is simple: your model, route, identity, and bill can all change before you get the answer. Buy through official channels, put spend alerts on AI accounts, and treat an API key like a company credit card.⁠ ⁠ AI security, Claude security, AI identity, API key security, token jacking, transfer stations, AI proxy, prompt routing, model verification, credential theft, AI supply chain, AI compliance, model routing, Anthropic, Unit 42, AI billing, agent governance, cloud security, audit logs, vendor risk⁠ ⁠ #Anthropic #ClaudeAI #AISecurity #Cybersecurity #APISecurity #ArtificialIntelligence⁠ by @cyphrexio
0
14 days ago
Download

Download All

Save the photos and videos currently loaded on this page.

0Photos
0Videos
Ready to download 0 / 0

Keep this tab open until your download is ready.
back to up