aramas 🪐 retweeted
shipping, shipping and a bit more shipping.
we saw teams struggling inside their own investigations, piecing together the data they need to fully understand a risk or alert. because of this, we’re adding more compounded security intelligence to Clarion’s Brain (contextualizing vulnerabilities and connecting the data scattered across environments).
the agentic swarm reasons across all findings, systems, software versions, etc., resulting in more investigative context for teams to act on in time.
Slack is fully integrated with Clarion. just create a thread to escalate the investigation, write a follow-up message, and Clarion will give you the full report.
check the full update: cantina.review/changelog-nov
aramas 🪐 retweeted
catching bugs that let bad actors on the same Wi-Fi change people’s code without ever touching their laptop.
Apex just found a high-severity bug in a tool that developers use to build, develop and test app interfaces. If the dev server was accessible over the network, someone on the same Wi-Fi could obtain its access credentials and use commands to inject malicious code into your project files.
not only that, but if the Vitest addon was installed, the injection could also lead to remote code execution on the device.
the Storybook team has fixed the issues. full writeup: cantina.review/storybook-3f8…
aramas 🪐 retweeted
Apex Flash hit more than 2.8k downloads in just over a week.
the future of security is ours to build. stay tuned as exciting news will be announced soon.
try it out: cantina.security/apex-flash
aramas 🪐 retweeted
Thrilled to finally get this out!
A few thoughts on the release:
1) The Fable 5 mania in June solidified the need for cybersecurity firms to own their destiny. It's not hard to envision a world where a certain political wind blows and models with advanced cyber capabilities are deemed contraband. You can either work towards accumulating the data and expertise you need to train your own or you can remain beholden to the labs' trusted access programs and API margins.
In the same vein, I don't think the regressions on cyber benchmarks for the 6/6.1 series of OAI models are a fluke - as mech interp style techniques get better, cyber/exploitation is on the shortlist of capabilities to be nerfed.
2) @shoucccc and his team at @yetalabs are extremely proficient and a pleasure to work with. Zero surprise there.
3) When starting out with Apex we would hit refusals and jailbreaks were the first thing we reached for.
I've spent a fair bit of time jailbreaking models and have found that there's a pretty significant "jailbreak tax"; models perform worse on the same task when in a jailbroken state. The Anthropic paper linked below imo directionally supports that conclusion in its content despite the title stating the inverse (many such cases). [1]
We found that jailbreaks:
a) are hard to craft and often rely on optimizing a specific prompt against a specific guardrail for a specific model
b) can be unreliable (a "good" jailbreak may still fail upwards of ~50% of the time on single turn tasks); you then have to try and recover back to a previous state before the model started clamming up
c) are complicated by long agentic tasks where the agent has ample opportunities to "remember" its guardrails
d) can leave the model in a deepfried state that might be able to bypass guardrails but has been pushed wildly out of distribution
e) don't add new capabilities
Given the data we accumulated from our bug bounty scanning with Apex, building training expertise became a far more appealing long term solution.
4) I find apex-flash-1 earns its keep as a bughunter grunt, typically orchestrated by a larger model. It fits in a lot of the same places as base 5.3f / ds 4.1f / luna but the rabbit holes it explores are different enough to warrant its use.
Enjoy!
---
[1] arxiv.org/pdf/2605.00267
today we're releasing apex-flash-1, our first open-weights model for security research post-trained on real vulnerabilities we found and got paid for.
we are also releasing the abliterated variant for researchers who want fewer refusals in their own authorized workflows.
how we built it: cantina.security/apex-flash
here we go!
today we're releasing apex-flash-1, our first open-weights model for security research post-trained on real vulnerabilities we found and got paid for.
we are also releasing the abliterated variant for researchers who want fewer refusals in their own authorized workflows.
how we built it: cantina.security/apex-flash
aramas 🪐 retweeted
we found multiple high severity vulnerabilities in an Australian open source server used to analyze population level healthcare and medicare data.
and there are many more to be found.
applications managing user health data need to adapt by regulation to a standard called FHIR (Fast Healthcare Interoperability Resource). in the United States, standard is mandated by federal regulation.
developers implement FHIR according to the specification to ensure patient data is well structured. but there's a catch, the compliance tests ONLY make sure that the standard is implemented CORRECTLY, not SECURELY.
there is a huge assurance gap in the industry right now affecting both public and private organizations.
we are actively cooperating with organizations to help them find these vulnerabilities before they go into production. if you are in the Health-Tech industry and implementing FHIR, reach out for a free vulnerability assessment.
Australia has been hacked.
'And today, I spoke with the CEO of OpenAI, Sam Altman, to express Australia's extreme concern about this incident. And I also expressed my disappointment that it took the company way too long to inform the government what had occurred, and the nature of the way that that notification occurred as well was unacceptable.'
no 1 feels good ✨
gg. we'll take the number 1 spot on @Hacker0x01's US leaderboard with our autonomous pentester. more than number 2, 3 and 4 combined.
We're giving our free FHIR vuln scans, to make hitting the 2027 CMS deadlines easier!
Your team can find out now whether the wrong person can reach patient records, instead of finding out when the stakes are real.
We pointed Apex at Pathling, an open source FHIR server used in hospitals.
It discovered 5 high-severity findings and 6 public CVEs, one allowed a token limited to a single patient to access records that it should never have been able to reach.
None of them would fail a conformance test.
Building FHIR APIs for 2027? We'll check yours free: cantina.review/free-fhir-62a…
aramas 🪐 retweeted
Clarion now reads Claude Code telemetry and turns every MCP server your agents touch into an audited and tracked asset.
For everything else new in the platform: cantina.review/claude-code-9…
aramas 🪐 retweeted
we spend a lot of time creating agents for security work but based on our conversations at blackhat not every security team does, so we are starting a new series sharing how we do it! hope it helps, give us feedback on what would you like to see more of!
cantina.review/web-exposure-…
aramas 🪐 retweeted
Answering the question everyone wants to know: we pit Apex against @claudeai and @OpenAI's Codex on the same codebase. Who comes out on top?
Of all the issues they detected, 78.6% of Apex’s findings were confirmed in the final review. For Claude, this number was 28.6%, and for Codex, just 16.7%.
The full method & results: cantina.review/claude-code-a…
aramas 🪐 retweeted
The next era of security will be a daily adversarial feedback loop.
In over 4,000 findings made by Apex, over a third were high or critical. Staying ahead of adversaries on a daily basis is the new mission.
Read our latest report: cantina.review/apex-field-1b…
aramas 🪐 retweeted
How can you find a bug in a compiler that has been mathematically proven to be correct?
Instead of focusing on the proof, you look for issues in everything else connected to it.
This is what our agentic OffSec engineer, Apex, did with CompCert, and it found 3 issues: cantina.review/bugs-ca2386
my face at black hat everytime I saw a claim about an agent doing something wildly cool but no proof that it actually did (versus @mydigitalself demoing exactly that value live on stage w one of cantina’s security agents)
ps how cool is our new brand
aramas 🪐 retweeted
Vendors publish their wins. @chrispyprojects, who taught Apex (our AI appsec solution) how to hunt, published its full scorecard against human audits, some costing $500,000+: every critical and high matched, plus live bugs the audits missed.
AI security claims should be backed by data. Here's ours, with something waiting there for your codebase: cantina.review/agentic-secur…
I think it’s very hard to defend that Web3/Blockchain Security Audits are not solved by Autonomous AI Bug Hunters like Cantina’s Apex.
Not only are we #1 on the HackerOne US business leaderboard, but we also (a very small team) used Apex to farm nearly $1M in bounties in the span of a few months, abusing the free money machine while competition was scarce.
We took every scan we’ve ever run (1500+) and looked carefully at the data and found some interesting trends we think are worth sharing with everyone.
1. # of valid bugs scales log-linearly with compute across 3 orders of magnitude. Double the compute and you get ~40% more findings on the median scan.
2. Scans saturate. There is a finite number of bugs in a codebase, and we can consistently predict once we’ve hit the ceiling of Apex’s capability to find more bugs in a codebase.
3. The indeterministic properties of AI bug hunters goes away as you scale compute. Apex more reliably converges on the same set of findings as you spend more compute.
4. The data supports that audits that would normally cost more than half a million dollars and take months, can be completed in a few days for a few thousand bucks, end-to-end by AI products. Solving the coverage problem took engineering an efficiently scaling harness. We are able to hit superhuman capabilities in our benchmarks for entirely automated security audits and the economics of it are ridiculous. For 1/100th, or in some case 1/1000th the cost, you get the same/better performance than a set of expert researchers.
Moving forward, we are seeing a new trend that is potentially scary. As more security researchers integrate AI more deeply into their workflow, at what point does hiring a researcher just mean you’re hiring their custom built harness? We’ve run scans side by side with recent audits comparing results after, and often see that the researcher findings are a subset of Apex findings. It’s very difficult now to measure the set of out-of-distributions bugs because of AI. Maybe manual hunting can make a comeback if your particular style is difficult to replace by AI.
Check out the full blog here: cantina.security/blog/ai-vs-…