Pinned Tweet
Finding the Firecracker bug was only the beginning. With @coinspect, we spent weeks proving a guest could run code in its host-side process, then chained a separate seccomp bypass. The escape is fixed (CVE-2026-122192). Full write-up:
coinspect.com/blog/firecrack…
JPMK retweeted
We reviewed Fluidkey's historical frontend changes for signs of malicious JavaScript or supply-chain compromise that could affect other projects.
We found no suspicious changes in the data we analyzed.
This is an independent review of public data, not part of the incident response. Await official updates.
JPMK retweeted
⚠️ Defenders are flooded with CVEs. Some are easier to exploit than others.
One worth checking: SigNoz CVE-2026-97055.
An unset JWT secret (default empty) lets attackers forge admin sessions.
If you run SigNoz:
→ Upgrade to 0.143.0+
→ Investigate suspicious access
→ Review what data your logs/traces leak
→ Rotate secrets as needed
A self-hosted instance was exploited in a crypto incident recently.
Hey everyone, I see twitter is on fire as usual.
We were impacted by a security vulnerability affecting SigNoz, resulting in approximately $200-250k in affected funds (less than 5% of total funds).
We're actively working on resolving the issue, and all affected users will be fully reimbursed. We'll publish all refunds throughout the day.
Sorry for the delay in communication, I was at an airport without internet access.
Please do not send any additional funds to the affected wallets.
The impact remains limited, and we're doing everything we can to resolve this as quickly as possible.
We'll keep you updated.
JPMK retweeted
Our analysis of Fluidkey’s v1.79.22 frontend changes is consistent with their explanation.
The old client could automatically sign backend-approved transactions without cryptographically binding approval to their contents.
The update introduces EIP-712 transaction-bound approvals.
We couldn't identify the session leak root cause from the frontend code, but the authorization changes match the attack they describe.
Fluidkey is back online with v1.79.22 on web, iOS and Android.
Previous versions will remain offline. To re-enable transactions on the app go to Settings > Update. We have implemented a block for all previous versions of the web and mobile app, so you will only be able to transact after the update to the safe new version.
The attacker used flaws in a third-party logging tool to obtain session data, prepare unauthorized transfers on our servers, and have them signed when a user opened the app. Transfers can no longer be prepared this way in v1.79.22, which also includes additional safeguards. The private keys controlling user funds never left the device.
Next, we will be reaching out to affected users individually and publish a full post-mortem.
JPMK retweeted
Testing browser agents like @googlechrome auto browse on the live web is tricky: sites change, side effects persist & testing prompt injection on live pages isn't safe!
So our team built a "Mirror Dimension": a hermetic, synthetic web to safely test agents! 🪞🤖
🧵
JPMK retweeted
Infosec likes to forget its history. Everyone should remember Apache-nosejob and how experts considered an Apache overflow unexploitable and then Gobbles just dropped the working exploit. Just because you think X is hard or unexploitable doesn’t mean another mind/AI can’t do it.
Congratulations to my coauthors for this great achievement.
Our paper “BABE: Verifying Proofs on Bitcoin Made 1000x Cheaper,” has won the Distinguished Paper Award at @acm_ccs 2026, one of the world's leading cybersecurity conferences.
Co-authored with researchers from UC Berkeley and Stanford, BABE is a core cryptographic engine behind Babylon TBV, enabling native Bitcoin to be used as trustless collateral without wrapping or bridging.
This recognition is a testament to the rigorous research and engineering behind making Bitcoin programmable preserving its trustless foundations.
JPMK retweeted
AI is changing everything fast. There's excitement about what's becoming possible, but also anxiety about jobs and seeing skills that took years to develop being automated.
In infosec some people are tired of exaggerated claims. Others are tired of risk being downplayed and patches delayed.
I think both sides have valid points, but these debates become polarized, leaving little room for opinions that don't fit neatly into either side.
Nice! we exploited this a different way in the original php2shell as mentioned in the writeup from quoted post.
almost every app uses session_start, session_decode, unserialize ..
for the original vuln this exploitation only requires an app use such a function (doesn't need controlled input) and groom profile for the given target but can remove the rest of the complicated memory corruption chain when a valid pop chain exists in the app (which is almost everything).
when there isnt a working POP chain, you can still RCE through the more complex memory corruption chain using native objects.
most apps dont focus on reducing exploitable pop chains though vs removing unserialize() for example.
"its not exploitable on servers with traffic", "it takes too much data to oom" -> skill issues..? which is crazy in the ai vr era.
some people figured out these tricks already but posting some slides from my @unpromptedconf talk (which I will upload later, though they are less good without me bumbling around on stage probably..)
PHP has just released a fix for CVE-2026-126714 (CVSS v4.0: 9.2), a critical vulnerability we discovered in the interaction between upload progress and PHP’s file-based session handler. If a pwrite()/ftruncate() update is interrupted, PHP may leave behind a malformed session file that is later loaded by session_start(), yielding an arbitrary deserialization primitive. Chained with our another PHP-native deserialization-to-RCE finding, this leads directly to remote code execution.
Overview: phpaftershell.com
Full technical details: depthfirst.com/research/phpa…
JPMK retweeted
‼️ UPDATE: The hackers who planted malware in Microsoft's AI agent toolkit on PyPI pushed it through Microsoft's own release system, with a stolen Microsoft certificate and private key.
Cloud security firm Wiz says their first try, from AWS, was blocked. The second, from a Microsoft-owned network, went through.
How the key was stolen is not yet known. Wiz counted three confirmed thefts of victim data before both versions were pulled within the hour.
The attackers' own AI hacking tool gave part of the operation away, reporting its progress to the same web address the stolen data went to.
JPMK retweeted
What happens when an agent interacts with a trusted but compromised website?
Prompt injection, or simply adding extra form fields or instructions to a page, turns the agent into a confused deputy: it holds your authority but takes direction from the attacker.
Social engineering for machines. And unlike (unlike?) humans, millions of agents share the same models, behaviors, and vulnerabilities, so attackers can test their attacks in advance.
JPMK retweeted
Some read my recent short posts about AI-developed exploits as dismissing AI-assisted vulnerability research, which isn’t my position. Short posts are a bad place for this topic, so here’s what I actually think:
Teams of experienced vulnerability researchers are using AI to find real vulnerabilities and develop working PoC exploits. I follow their work closely, use these tools myself, and I'm impressed.
A demo video can be the visible result of weeks of experiments and skillful guidance of AI agents. That work deserves recognition.
But that video is often all most people see and context gets lost: the environment, the assumptions, the limitations, and how much of the development was actually autonomous. We can end up arguing about different things without realizing it.
We may be assuming claims the researchers never made.
My skepticism is about one specific claim: that an inexperienced operator can hand a generic agent a simple goal and get back a reliable exploit against real production targets.
Years ago, I worked on what we called commercial-grade exploits. These weren't just demonstrations. They were used to identify vulnerable assets without false positives, exploit them reliably across real production environments, and minimize collateral damage. That meant testing different versions, platforms, and configurations, dealing with memory layouts and mitigations, and endless trial and error and QA.
That’s a different problem from exploitation in a controlled lab, and I haven’t yet seen evidence that AI solves it without experienced guidance, tools and testing infrastructure. This is what some call "weaponizing"
I may be wrong about how fast that changes. I spent years learning how hard reliable exploitation is. If AI is overcoming these challenges, I want to be among the first to know, and I’ll say so publicly. If you have evidence you can share, I want to see it.
None of this is a reason to dismiss vulnerabilities.
A lab PoC from a credible team is meaningful evidence.
Evaluate your exposure and patch. Don’t wait for a polished attack.
The culture of downplaying vulnerabilities and delaying patches needs to change.
I hope we can challenge claims without attacking people, disagree without dismissing each other's experience or achievements. I’ll try to do a better job explaining my own views more clearly.
JPMK retweeted
Great research by @BitFinding
No exotic smart contract bug.
Just (bit) finding the gap between what a system considers confirmed and what can still be reversed.
Without the nonce trick, the deposit tx could simply be included again, allowing the relayer to recover its funds.
Orphan blocks can be used to drain bridge relayers. Here’s how we abused them to steal funds when Ethereum reorged.
bitfinding.com/blog/the-orph…
JPMK retweeted
Great to see the Ethereum Foundation pledging $27K to support DappFence through TheDAO Security Fund!
DappFence is a free, open-source solution built on web standards that lets users verify the JavaScript they run, without having to trust the Web2 infrastructure delivering it.
JPMK retweeted
What happens when a bridge pays out before the source-chain deposit is final, and that deposit disappears?
Great research by @Bitfinding, demonstrating this against Across, Relay and LI FI relayers.
If you're building or auditing cross-chain bridges, this is a must-read.
Orphan blocks can be used to drain bridge relayers. Here’s how we abused them to steal funds when Ethereum reorged.
bitfinding.com/blog/the-orph…
JPMK retweeted
We received an alert about obfuscated code interacting with wallets on the 1inch frontend.
Looks like malware, but it is Innerworks SDK, a security provider.
The code reads wallet addresses and sends them, with a persistent user ID and device fingerprint, to Innerworks.
JPMK retweeted
Orphan blocks can be used to drain bridge relayers. Here’s how we abused them to steal funds when Ethereum reorged.
bitfinding.com/blog/the-orph…
Since the vulnerability has leaked: Calif discovered and reported Java4Shell (CVE-2026-114938).
We're publishing technical details by 1 PM Pacific today, after Oracle releases a patch.
Stay tuned!
JPMK retweeted
Replying to @rsquare
Did it really take Safe a full year to remove Beamer after we warned about the risks of loading unnecessary third-party JavaScript?



























