From my experience all software developers are now security engineers wether they know it, admit to it or do it. Your code is now the security of the org you work for. #GoldenAgeOfDefense

IR folks (@GelosSnake @ProferoSec @Mandiant @Volexity @attrc @uuallan @TalosSecurity others, please tag): trying to put numbers on breaches. Vs. a year ago (or whatever makes sense), how many more cases are you seeing? Are you really "turning away work" as @RGB_Lights says?

Calif was founded with a mission to make the Internet a more secure place for everyone. In the last few months, we have been seeing significant advances in AI model capabilities, especially in the domain of offensive security research. In addition to publishing our research and learnings, we wondered what else we could do to help defenders. While large corporations and government agencies have the dedicated budget and human resources for incident response and forensics, there are non-profit organizations –- collectively known as civil society organizations –- investigating digital threats affecting vulnerable groups and they are funded entirely by volunteers. In order to perform their job, the civil society researchers may need to extract data from mobile devices for forensic purposes, reverse engineer malware, and conduct similar security tasks. This work can be challenging given the restricted environments on some mobile devices. We understand these technical challenges and will offer our research services and capabilities to these organizations, free of charge. We support them and their mission. If you represent such an organization and need assistance, contact us at help@calif.io.

I appreciate this balanced and realistic perspective.

Stop Russian compromise of your network devices by hardening your SNMP. We urge NSS and DIB admins to move to SNMPv3 immediately to ensure payloads are authenticated and encrypted. Full hardening guidance here: nsa.gov/Cybersecurity/Cybers…

Many companies are waking up to the realization that basic cybersecurity hygiene was actually important after all. Or maybe they thought they already had it. These attacks aren’t exploiting 0days or highly complex vulns. We have the controls to protect against these attacks already, today. I’ll say it again: the era of leaving your low-hanging fruit unpicked is over. However, security companies should be ashamed when they reword statements to stoke fear and make the situation appear like something it isn’t. I guess levels of shame don’t show up in QBRs. cert-in.org.in/s2cMainServle…

Brilliant perspective! 🤙

Right now slop isn’t really the problem. Programs budgeted for a trickle of P1s and they’re getting a hose pipe of genuine crits. They're out of budget/time or will be soon, and they can’t fix bugs faster. So they are reducing the amount of people who submit instead.

There are two very frustrating things going on with defenders today. First, encountering critical infrastructure that obviously hasn’t been hardened by AI at all. Second, defenders thinking “48 hours to a vuln response” is a timeline they can afford.

Transition from vulpocalypse to breachpocalypse is here. IR firms are saturated & turning away work. Now one hacker unleashed AI agents on 1,200+ websites, compromising 100+ for ~$5,000 in tokens. $50 per hack. Agents won and compromised targets. reuters.com/legal/government…

I would NEVER use an @AnthropicAI or @OpenAI model for security purposes. If you’re relying on this (daybreak or no daybreak) you have a critical point of failure in your enterprise IR. This has been proven in multiple breach cases now. Ditch them.

Context about how I got CVE-2026-99999: it wasn't assigned by incrimental luck. There were nearly 2k unassigned CVEs before it. The Linux maintainers felt it deserved the special number because of impact+new attack surface for KVM escapes and offered it. It was a prize! I was just telling someone it feels like graduation from cybersecurity. So yeah it's intentionally assigned, and I'll be eternally grateful for it.

Talking AI and secure coding across Germany! 🇩🇪 So far: Hamburg, Hanover, Bochum, Cologne and Stuttgart. Today: Karlsruhe! Up next: Oct 21: Frankfurt Oct 23: Augsburg Oct 26: Munich Thank you to everyone hosting and joining me!

This is the single greatest DNS post of all time. 😎

My formula for how many AI sessions I can run at once: A=⌊(2+3ln(1+C))(1−e^(−S/2))/((1+e^(4−S))(1+(max(0,C−3)/2)^4))⌋ C = coffees. S = hours slept. At 6 coffees and 3 hours of sleep, the agents open a session to manage me. Peer-reviewed by my espresso machine. #ManicodeMath 😁

This is an awesome thread from my favorite DNSSEC wizard. 😎

This program is a pain and complex. I had similar problems but went through support. Here’s a very long post explaining how to troubleshoot. I'm seeing three block scenarios. 1) You were accepted into the CVP, but the approval isn't applied to the account you're actually using. CVP approval is tied to a specific organization ID, not to you as a person. If you're working in your personal workspace instead of the org that was approved, you get default safeguards. Compare the org ID where you're seeing blocks with the one on your approval email. For API and Console-based access, approval arrives as a program (grant) on your org, and it does nothing until an org Admin or Owner assigns it to a workspace under Settings > Programs and sets it to Active. Some programs can't be assigned to the default workspace, so you may need a dedicated one. Check the workspace's Qualifications panel for unmet requirements. Some programs also have seat caps, and if the org goes over, assigned workspaces lose access until you're back under the limit. On Enterprise plans the program is assigned through custom roles, so make sure your users are actually in that role. Also check how you're reaching Claude. Bedrock support is very limited right now, Google Cloud requires data retention to be configured for every project, region, and model combination, and third-party tools only get CVP if that platform participates. The program generally requires data retention, so Zero Data Retention orgs can be excluded. 2) You have Defense Access, but you're doing offensive work. The tiers are scoped to the work you're authorized to do. Defense Access covers defending systems your org owns or maintains: SOC and incident response, malware analysis, and vulnerability validation. Pen testing and red teaming require Red Team Access. Anthropic's own testing shows how big the gap is. On their CyScenarioBench evaluation of multistage cyber operations, Defense Access blocked 46 of 50 trials. Red Team Access blocked none. If you're running an engagement on Defense Access, the blocks are expected behavior. Two things catch people here. Red Team Access is for organizations only, so individual applicants can only get Defense. And if your org applied for Red Team, you're enrolled in Defense while that review runs, which takes weeks. Plenty of people who think they have Red Team access are still on Defense. 3) You legitimately have Red Team Access and still get blocked. The usual causes: - Prohibited use is blocked at every tier. Ransomware deployment, mass data exfiltration, and actions that could harm people or cause widespread disruption stay blocked no matter what. - Wrong model. The CVP covers Opus 5.5, Sonnet 5.5, and Mythos 5.1. Fable 5.1 is not a part of the CVP. - Wrong workspace or credential. Red Team Access comes with business verification, phishing-resistant authentication, and credential and seat controls. The program only applies to traffic from workspaces that meet its requirements, so a script using a credential scoped to another workspace gets default behavior. - The classifiers aren't perfect. Anthropic says plainly that approved users may still hit blocks on legitimate work. If you've ruled all of that out, file a report or appeal here: claude.com/form/cyber-block-… Include the org ID, workspace, model, and roughly when it happened. That data is what they use to tune the classifiers. Useful references: - Cyber safeguards and CVP: support.claude.com/en/articl… - Assigning a program to a workspace: support.claude.com/en/articl…

A lot of my good friends are struggling with the Anthropic CVP program. The OpenAI CVP program seems really smooth. This is only a problem with Anthropic. Here is a troubleshoot guide to help.

AI is testing the limits of post-quantum cryptography. An Anthropic AI model found an attack that halved HAWK's effective security in just 60 hours, prompting its withdrawal from NIST consideration. Itan Barmes of QIZ Security explains why cryptographic agility now matters more than ever. Read: thehackernews.com/expert-ins…

The #vulncopalyse is real

dear future ai harness builders, we don't need another ai harness thanks, can