Vinu Thomas (@vinuthomas) header
Vinu Thomas (@vinuthomas) profile picture

Music, Photography, Tech, Security

Bangalore
Joined March 2007
Photo by Vinu Thomas (@vinuthomas)Photo by Vinu Thomas (@vinuthomas)Photo by Vinu Thomas (@vinuthomas)Photo by Vinu Thomas (@vinuthomas)Photo by Vinu Thomas (@vinuthomas)Photo by Vinu Thomas (@vinuthomas)Photo by Vinu Thomas (@vinuthomas)Photo by Vinu Thomas (@vinuthomas)Photo by Vinu Thomas (@vinuthomas)Photo by Vinu Thomas (@vinuthomas)
When you configure OpenSSH on an Alpine Linux server, the way password authentication is handled can depend on whether the server uses PAM and how the underlying C library handles password hashing. The issue affects specific configurations: - OpenSSH server built without PAM - OpenSSH server with PAM installed but `UsePAM no` configured - Password authentication enabled through `PasswordAuthentication yes` OpenSSH runs a dummy password check to make it harder to tell whether an account exists by timing the serverโ€™s response. On affected Alpine systems, OpenSSH could treat an error from the password hashing library as a successful match. That turned a timing defence into an authentication bypass. If you run SSH on Alpine, update to the patched OpenSSH package for your Alpine release and restart `sshd`. The package update alone leaves the old daemon running. If you canโ€™t patch immediately, confirm you can sign in with an SSH key, then set `PasswordAuthentication no` and restart `sshd` This case also raises a broader engineering question: how carefully do we handle error values when security-critical components depend on the behaviour of underlying libraries? #Cybersecurity #OpenSSH #Linux #SoftwareEngineering
42
A busy day can hit your database twice. ๐—ข๐—ป๐—ฐ๐—ฒ ๐˜„๐—ต๐—ฒ๐—ป ๐˜๐—ต๐—ฒ ๐˜„๐—ผ๐—ฟ๐—ธ ๐—ฟ๐˜‚๐—ป๐˜€ ๐—ฎ๐—ป๐—ฑ ๐—ฎ๐—ด๐—ฎ๐—ถ๐—ป ๐˜„๐—ต๐—ฒ๐—ป ๐˜†๐—ผ๐˜‚ ๐—ฑ๐—ฒ๐—น๐—ฒ๐˜๐—ฒ ๐—ถ๐˜. Thatโ€™s what caught my attention in CircleCIโ€™s October 9 incident report. CircleCI used Temporal to coordinate work, with PostgreSQL storing its workflow history. Completed workflows were deleted after three days. When those deletions kicked in, they generated more write-ahead log (WAL) activity than the database could archive. The dedicated log volume filled up and PostgreSQL stopped accepting writes. Moving the transaction log to a volume with spare capacity allowed PostgreSQL to accept writes again, but another bottleneck emerged. Millions of deleted rows were still awaiting cleanup. ๐—ง๐—ต๐—ถ๐˜€ ๐—ถ๐—ป๐—ฐ๐—ฟ๐—ฒ๐—ฎ๐˜€๐—ฒ๐—ฑ ๐˜๐—ต๐—ฒ ๐˜„๐—ผ๐—ฟ๐—ธ ๐—ฃ๐—ผ๐˜€๐˜๐—ด๐—ฟ๐—ฒ๐—ฆ๐—ค๐—Ÿ ๐—ต๐—ฎ๐—ฑ ๐˜๐—ผ ๐—ฑ๐—ผ, ๐—ฑ๐—ฟ๐—ถ๐˜ƒ๐—ถ๐—ป๐—ด ๐—–๐—ฃ๐—จ ๐˜‚๐˜€๐—ฎ๐—ด๐—ฒ ๐—ต๐—ถ๐—ด๐—ต๐—ฒ๐—ฟ. Attempts to speed up cleanup competed with live queries for the same capacity, while upstream services kept submitting new work and retrying failures. The bottleneck had shifted from log storage to CPU, and the retries made it worse. For me, the useful lesson is that ๐—ฟ๐—ฒ๐˜๐—ฒ๐—ป๐˜๐—ถ๐—ผ๐—ป ๐—ฝ๐—ผ๐—น๐—ถ๐—ฐ๐—ถ๐—ฒ๐˜€ ๐—ฎ๐—ฟ๐—ฒ ๐—ฎ๐—น๐˜€๐—ผ ๐—ฐ๐—ฎ๐—ฝ๐—ฎ๐—ฐ๐—ถ๐˜๐˜† ๐—ฑ๐—ฒ๐—ฐ๐—ถ๐˜€๐—ถ๐—ผ๐—ป๐˜€. Choosing how long to keep data determines when your infrastructure has to do the work of removing it. In a MarTech platform, ๐˜†๐—ฒ๐˜€๐˜๐—ฒ๐—ฟ๐—ฑ๐—ฎ๐˜†โ€™๐˜€ ๐—ฟ๐—ฒ๐˜๐—ฒ๐—ป๐˜๐—ถ๐—ผ๐—ป ๐—ฑ๐—ฒ๐—น๐—ฒ๐˜๐—ฒ๐˜€ ๐—ฐ๐—ผ๐˜‚๐—น๐—ฑ ๐—ฐ๐—ผ๐—บ๐—ฝ๐—ฒ๐˜๐—ฒ ๐˜„๐—ถ๐˜๐—ต ๐˜๐—ผ๐—ฑ๐—ฎ๐˜†โ€™๐˜€ ๐—ฎ๐˜‚๐—ฑ๐—ถ๐—ฒ๐—ป๐—ฐ๐—ฒ ๐—ฟ๐—ฒ๐—ณ๐—ฟ๐—ฒ๐˜€๐—ต๐—ฒ๐˜€ ๐—ฎ๐—ป๐—ฑ ๐—น๐—ถ๐˜ƒ๐—ฒ ๐—ฐ๐˜‚๐˜€๐˜๐—ผ๐—บ๐—ฒ๐—ฟ ๐—พ๐˜‚๐—ฒ๐—ฟ๐—ถ๐—ฒ๐˜€. When evaluating a platformโ€™s performance, Iโ€™d want to know: - How much database load can cleanup jobs generate? - Can we throttle them when live traffic needs capacity? - Do retries back off when a dependency is struggling? - Are we monitoring cleanup lag and ๐—ช๐—”๐—Ÿ archiving before customers notice delays? CircleCI eventually paused incoming and retry work to let the database recover, then gradually increased the load. ๐—•๐—ฎ๐—ฐ๐—ธ๐—ด๐—ฟ๐—ผ๐˜‚๐—ป๐—ฑ ๐˜„๐—ผ๐—ฟ๐—ธ ๐—ป๐—ฒ๐—ฒ๐—ฑ๐˜€ ๐—ฎ ๐—ฝ๐—ฟ๐—ผ๐—ฑ๐˜‚๐—ฐ๐˜๐—ถ๐—ผ๐—ป ๐—ฏ๐˜‚๐—ฑ๐—ด๐—ฒ๐˜ ๐˜๐—ผ๐—ผ. Which cleanup job in your system could disrupt live traffic? [CircleCIโ€™s postmortem](discuss.circleci.com/t/post-โ€ฆ) \#DatabaseEngineering #SystemReliability #MarTech #PerformanceEngineering
3
3
246
BBMP dug up our footpaths in Cooke Town. Somehow, we ended up paying for all the electrical repair and losing appliances. For the past month, BBMP has been digging up the footpaths and drains in our area. Thereโ€™s barely anywhere to walk except on the road competing for space between vehicles. Progress has been painfully slow. Then the digging damaged the power line to our apartment building. The damaged section was cemented over by the workers without anyone telling us. We started getting power fluctuations and I spent hours running up and down the aparment first, then the building trying to work out what was wrong. BESCOM confirmed that the cable had been damaged during the digging. Their solution? We had to pay for a replacement line into the building. We paid for the work and materials, expecting that would resolve it. Instead, the repair left us with even more fluctuations, voltage drops and power outage across the building. When we called again, we were told everything was fine on their end. It took an electrician we hired separately to identify the faulty work and point it out to BESCOM before they came back and fixed it today. By then, the damage included our washing machine, smart plugs and UPS, plus microwaves and induction cooktops in other flats. So weโ€™ve paid BESCOM for the repair, paid an electrician for repeated visits and now have damaged appliances to deal with. Thatโ€™s on top of the time and patience spent getting a problem fixed that we didnโ€™t cause. Who takes responsibility for the damage and the costs residents have been left with? As usual citizens have to pay for the complancy of lax goverment entities.
162
Built a Bluetooth shutter remote for the Insta360 GO 3S using a Flipper Zero. Photo capture and video start/stop are working. Required a small firmware patch to get BLE discovery working. Open source. Other Insta360 models need testing. github.com/vinuthomas/flippeโ€ฆ #FlipperZero #Insta360
1
1
194
Every year, I find myself reminding my team of one thing: ๐ญ๐ก๐ž๐ซ๐ž'๐ฌ ๐ฉ๐ฅ๐ž๐ง๐ญ๐ฒ ๐จ๐Ÿ ๐ญ๐ข๐ฆ๐ž ๐ญ๐จ ๐ ๐ž๐ญ ๐›๐š๐œ๐ค ๐ญ๐จ ๐›๐ฎ๐ข๐ฅ๐๐ข๐ง๐ . Sometimes, we just need to step back and remember that there's just as much to life outside work as there is in it. It's easy to get caught up in deadlines, to-do lists and the feeling that there's always something more to get done. But we all deserve a little time to ๐ฌ๐ญ๐ž๐ฉ ๐š๐ฐ๐š๐ฒ, ๐ฌ๐ฅ๐จ๐ฐ ๐๐จ๐ฐ๐ง ๐š๐ง๐ ๐ž๐ง๐ฃ๐จ๐ฒ ๐ญ๐ก๐ž ๐ฆ๐จ๐ฆ๐ž๐ง๐ญ without worrying about what's next. Diwali is a lovely opportunity to do just that. Whether you're spending time with family and friends, travelling somewhere new, catching up with old friends or simply enjoying a quiet day, I hope you make time for the things that make you happy. Wishing everyone a very happy Diwali! ๐Ÿช” P.S. The work can wait. Make the most of the celebrations!
2
91
What if we could ๐ฎ๐ง๐๐ž๐ซ๐ฌ๐ญ๐š๐ง๐ ๐ฅ๐ž๐ ๐š๐œ๐ฒ ๐ฌ๐จ๐Ÿ๐ญ๐ฐ๐š๐ซ๐ž without relying on its original source code? AI-assisted reverse engineering is creating new opportunities for software modernisation. Engineers can investigate how existing applications work, recover undocumented logic and test reconstructed code against the original implementation. The real value lies in what this makes possible: โ€ข Recover business rules hidden in legacy code. โ€ข Trace dependencies to understand how different components interact. โ€ข Verify reconstructed functionality before deciding what to replace or rebuild. Projects like DX-Ball demonstrate how engineers can recover and validate code against an existing executable. Applying these techniques to enterprise software opens up further possibilities, from investigating undocumented systems to making more informed modernisation decisions. The opportunity isn't just to understand old software. It's to make modernisation ๐ฆ๐จ๐ซ๐ž ๐ฉ๐ซ๐ž๐œ๐ข๐ฌ๐ž, ๐ญ๐š๐ซ๐ ๐ž๐ญ๐ž๐ ๐š๐ง๐ ๐ฏ๐ž๐ซ๐ข๐Ÿ๐ข๐š๐›๐ฅ๐ž. In my latest article, I explore how AI-assisted reverse engineering works, what current projects demonstrate and where the approach can make a difference. vinuthomas.com/reverse-enginโ€ฆ #AI #rea #SoftwareModernisation #ReverseEngineering #SoftwareEngineering
2
98
โ€ผ๏ธWARNING - OpenSSH issues emergency fix for a critical SSH authentication bypass on Alpine Linux. A password longer than 256 bytes can unlock key-only accounts on affected servers, without knowing the actual credentials. Here's how it works โ†“ thehackernews.com/2026/11/opโ€ฆ
88
665
232
3,160
911,348
A critical vulnerability lands. What needs fixing? Cloudflareโ€™s Zero Wire offers short briefings on whoโ€™s affected, whatโ€™s confirmed and what to do next, with links to the evidence. radar.cloudflare.com/securitโ€ฆ
62
Two iconic bridges, one timeless river. Evenings by the Hooghly never get old. #KolkataDiaries #VidyasagarSetu #CityOfJoy #prinsepghat #travel
81
๐—ข๐—ป๐—ฒ ๐˜๐—ต๐—ถ๐—ป๐—ด ๐˜๐—ต๐—ฎ๐˜ ๐—ฐ๐—ฎ๐—ป ๐—ฐ๐—ฎ๐˜๐—ฐ๐—ต ๐˜‚๐˜€ ๐—ผ๐—ณ๐—ณ ๐—ด๐˜‚๐—ฎ๐—ฟ๐—ฑ ๐—ถ๐—ป ๐˜๐—ฒ๐—ฐ๐—ต๐—ป๐—ผ๐—น๐—ผ๐—ด๐˜† ๐—ฝ๐—ฟ๐—ผ๐—ท๐—ฒ๐—ฐ๐˜๐˜€ ๐—ถ๐˜€ ๐—ต๐—ผ๐˜„ ๐—บ๐˜‚๐—ฐ๐—ต ๐˜„๐—ผ๐—ฟ๐—ธ ๐—ฟ๐—ฒ๐—บ๐—ฎ๐—ถ๐—ป๐˜€ ๐—ฎ๐—ณ๐˜๐—ฒ๐—ฟ ๐˜๐—ต๐—ฒ ๐—ณ๐—ฒ๐—ฎ๐˜๐˜‚๐—ฟ๐—ฒ ๐—ถ๐˜๐˜€๐—ฒ๐—น๐—ณ ๐—ถ๐˜€ ๐—ฏ๐˜‚๐—ถ๐—น๐˜. Imagine an AI assistant that helps employees resolve IT issues like password resets and VPN access. Building the assistant is one part of the job. Connecting it to internal knowledge bases and ticketing systems, managing permissions and ensuring it gives reliable answers are separate challenges. Then comes the cost of running it. If usage grows but the assistant fails to resolve enough queries, the company could end up paying for the AI service while still carrying most of its original support workload. ๐—” ๐˜„๐—ผ๐—ฟ๐—ธ๐—ถ๐—ป๐—ด ๐—ณ๐—ฒ๐—ฎ๐˜๐˜‚๐—ฟ๐—ฒ ๐—ถ๐˜€ ๐—ป๐—ผ๐˜ ๐—ฒ๐—ป๐—ผ๐˜‚๐—ด๐—ต ๐˜๐—ผ ๐—ท๐˜‚๐˜€๐˜๐—ถ๐—ณ๐˜† ๐˜๐—ต๐—ฒ ๐—ถ๐—ป๐˜ƒ๐—ฒ๐˜€๐˜๐—บ๐—ฒ๐—ป๐˜. The solution needs to deliver measurable value at a sustainable cost. Before committing to a wider rollout, Iโ€™d focus on three questions: ๐—ง๐—ฒ๐—ฐ๐—ต๐—ป๐—ถ๐—ฐ๐—ฎ๐—น ๐—ฑ๐—ฒ๐—ฝ๐—ฒ๐—ป๐—ฑ๐—ฒ๐—ป๐—ฐ๐—ถ๐—ฒ๐˜€: Are the integrations, data quality and access controls reliable enough for production? ๐—–๐—ผ๐˜€๐˜ ๐—ฝ๐—ฒ๐—ฟ ๐—ผ๐˜‚๐˜๐—ฐ๐—ผ๐—บ๐—ฒ: What does each successfully resolved request cost, and how does that change as usage grows? ๐—˜๐˜ƒ๐—ถ๐—ฑ๐—ฒ๐—ป๐—ฐ๐—ฒ ๐—ณ๐—ผ๐—ฟ ๐˜€๐—ฐ๐—ฎ๐—น๐—ฒ: Is the assistant reducing the support team's workload enough to justify expanding its use? If the numbers don't hold up, the answer may be to narrow the use case, address the underlying constraints or rethink the investment. ๐—ง๐—ต๐—ฒ ๐—ฟ๐—ฒ๐—ฎ๐—น ๐˜๐—ฒ๐˜€๐˜ ๐—ถ๐˜€๐—ป'๐˜ ๐˜„๐—ต๐—ฒ๐˜๐—ต๐—ฒ๐—ฟ ๐˜„๐—ฒ ๐—ฐ๐—ฎ๐—ป ๐—ฏ๐˜‚๐—ถ๐—น๐—ฑ ๐˜๐—ต๐—ฒ ๐˜๐—ฒ๐—ฐ๐—ต๐—ป๐—ผ๐—น๐—ผ๐—ด๐˜†. ๐—œ๐˜'๐˜€ ๐˜„๐—ต๐—ฒ๐˜๐—ต๐—ฒ๐—ฟ ๐˜๐—ต๐—ฒ ๐—ฟ๐—ฒ๐˜€๐˜‚๐—น๐˜๐˜€ ๐—ท๐˜‚๐˜€๐˜๐—ถ๐—ณ๐˜† ๐˜๐—ฎ๐—ธ๐—ถ๐—ป๐—ด ๐—ถ๐˜ ๐—ณ๐˜‚๐—ฟ๐˜๐—ต๐—ฒ๐—ฟ. How do you decide when a promising technology project is ready to scale? #AIImplementation #TechnologyStrategy #EnterpriseAI #ITLeadership
1
3
177
๐—” ๐—ณ๐—ฎ๐˜€๐˜๐—ฒ๐—ฟ ๐—–๐—ฃ๐—จ ๐—ฑ๐—ผ๐—ฒ๐˜€๐—ป'๐˜ ๐—ฎ๐—น๐˜„๐—ฎ๐˜†๐˜€ ๐—บ๐—ฒ๐—ฎ๐—ป ๐—ฎ ๐—ฐ๐—ต๐—ฒ๐—ฎ๐—ฝ๐—ฒ๐—ฟ ๐—ฑ๐—ฎ๐˜๐—ฎ ๐—ฐ๐—ฒ๐—ป๐˜๐—ฟ๐—ฒ. When you're running thousands of servers, even a small difference in power consumption can add up. And a processor that performs well in benchmarks may not deliver the same advantage under everyday workloads. This is part of what makes Nuvacore interesting. The Sequoia-backed startup is developing CPU cores for data centres and is reportedly raising funds at a valuation of around $2.5 billion. Nuvacore is taking a different route with its ๐—–๐—ผ๐—ฟ๐—ฒ ๐—™๐—ถ๐—ฟ๐˜€๐˜ ๐—ฎ๐—ฝ๐—ฝ๐—ฟ๐—ผ๐—ฎ๐—ฐ๐—ต. Rather than choosing an instruction-set architecture like Arm or x86 upfront, it's developing the processor core first. The thinking is that this gives engineers more room to work on performance and efficiency before settling on the architecture. There's a trade-off, though. ๐—ง๐—ต๐—ฒ ๐—ถ๐—ป๐˜€๐˜๐—ฟ๐˜‚๐—ฐ๐˜๐—ถ๐—ผ๐—ป-๐˜€๐—ฒ๐˜ ๐—ฎ๐—ฟ๐—ฐ๐—ต๐—ถ๐˜๐—ฒ๐—ฐ๐˜๐˜‚๐—ฟ๐—ฒ ๐—ฎ๐—ณ๐—ณ๐—ฒ๐—ฐ๐˜๐˜€ ๐˜„๐—ต๐—ถ๐—ฐ๐—ต ๐˜€๐—ผ๐—ณ๐˜๐˜„๐—ฎ๐—ฟ๐—ฒ ๐—ฎ ๐—ฝ๐—ฟ๐—ผ๐—ฐ๐—ฒ๐˜€๐˜€๐—ผ๐—ฟ ๐—ฐ๐—ฎ๐—ป ๐—ฟ๐˜‚๐—ป. A better core won't help much if adopting it means expensive changes to the applications and infrastructure already in place. If I were evaluating a new CPU for a data centre, I'd want answers to three questions: โ†’ ๐—ฃ๐—ฒ๐—ฟ๐—ณ๐—ผ๐—ฟ๐—บ๐—ฎ๐—ป๐—ฐ๐—ฒ: Do the gains hold up under the workloads we actually run? โ†’ ๐—ฃ๐—ผ๐˜„๐—ฒ๐—ฟ: How much useful work are we getting for the electricity consumed? โ†’ ๐—”๐—ฑ๐—ผ๐—ฝ๐˜๐—ถ๐—ผ๐—ป: Will the savings justify the cost of making our existing software and systems work with it? That's why I wouldn't judge a new processor on benchmark numbers alone. ๐—ง๐—ต๐—ฒ ๐˜€๐—ผ๐—ณ๐˜๐˜„๐—ฎ๐—ฟ๐—ฒ ๐—ฒ๐—ฐ๐—ผ๐˜€๐˜†๐˜€๐˜๐—ฒ๐—บ ๐—ฎ๐—ป๐—ฑ ๐˜๐—ต๐—ฒ ๐—ฐ๐—ผ๐˜€๐˜ ๐—ผ๐—ณ ๐—ฏ๐—ฟ๐—ถ๐—ป๐—ด๐—ถ๐—ป๐—ด ๐—ถ๐˜ ๐—ถ๐—ป๐˜๐—ผ ๐—ฝ๐—ฟ๐—ผ๐—ฑ๐˜‚๐—ฐ๐˜๐—ถ๐—ผ๐—ป ๐—บ๐—ฎ๐˜๐˜๐—ฒ๐—ฟ ๐˜๐—ผ๐—ผ. @Nuvacoreai still has to demonstrate what its approach can deliver. For now, the design choice is interesting, but the results will determine whether it's practical at data-centre scale. Would you consider moving to a new CPU architecture for better efficiency if it meant changing parts of your existing software stack? Source: reuters.com/legal/transactioโ€ฆ #Semiconductors #DataCenters #CPUDesign #DataCenterEfficiency #ProcessorDesign
101
This is Genie 4, our most realistic and interactive world model yet. ๐ŸŒ Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.
1,029
2,636
2,055
23,820
3,554,686
When the sky wakes up in flames and the geese welcome in a brand-new day. Morning magic over Lake Simcoe, Barrie. ๐Ÿ‡จ๐Ÿ‡ฆ #MorningGlow #SkyOnFire #travelmoments #barrieontario
1
1
2
138
๐—ช๐—ฒโ€™๐˜ƒ๐—ฒ ๐˜€๐—ฝ๐—ฒ๐—ป๐˜ ๐—ฑ๐—ฒ๐—ฐ๐—ฎ๐—ฑ๐—ฒ๐˜€ ๐—ฑ๐—ฒ๐˜€๐—ถ๐—ด๐—ป๐—ถ๐—ป๐—ด ๐˜€๐˜†๐˜€๐˜๐—ฒ๐—บ๐˜€ ๐—ฎ๐—ฟ๐—ผ๐˜‚๐—ป๐—ฑ ๐—ฎ ๐˜€๐—ถ๐—บ๐—ฝ๐—น๐—ฒ ๐—ฎ๐˜€๐˜€๐˜‚๐—บ๐—ฝ๐˜๐—ถ๐—ผ๐—ป: ๐—ถ๐—ณ ๐˜€๐—ผ๐—บ๐—ฒ๐˜๐—ต๐—ถ๐—ป๐—ด ๐—ด๐—ผ๐—ฒ๐˜€ ๐˜„๐—ฟ๐—ผ๐—ป๐—ด, ๐˜„๐—ฒ ๐—ฐ๐—ฎ๐—ป ๐˜๐—ฟ๐—ฎ๐—ฐ๐—ฒ ๐˜๐—ต๐—ฒ ๐—ฑ๐—ฒ๐—ฐ๐—ถ๐˜€๐—ถ๐—ผ๐—ป. Agentic AI is starting to break that assumption. AI is moving from executing defined instructions to interpreting intent, choosing actions and interacting with systems on our behalf. Weโ€™re already seeing the cracks: โ€ข AI can confidently make the wrong call while users treat it as authoritative. โ€ข Agents can take actions their creators or users never explicitly intended. The obvious question is: ๐˜„๐—ต๐—ผ ๐—ถ๐˜€ ๐—น๐—ถ๐—ฎ๐—ฏ๐—น๐—ฒ? The more important question for technology leaders is: ๐—ฐ๐—ฎ๐—ป ๐—ผ๐˜‚๐—ฟ ๐˜€๐˜†๐˜€๐˜๐—ฒ๐—บ๐˜€ ๐—ฒ๐˜ƒ๐—ฒ๐—ป ๐—ฒ๐˜…๐—ฝ๐—น๐—ฎ๐—ถ๐—ป ๐—ต๐—ผ๐˜„ ๐˜๐—ต๐—ฒ ๐—ฑ๐—ฒ๐—ฐ๐—ถ๐˜€๐—ถ๐—ผ๐—ป ๐—ต๐—ฎ๐—ฝ๐—ฝ๐—ฒ๐—ป๐—ฒ๐—ฑ? That changes how we think about architecture. Audit trails. Permissions. Human checkpoints. Failure boundaries. Observability. These used to be engineering safeguards. With autonomous systems, they become part of the product itself. ๐—œ๐—ณ ๐—ฎ๐—ป ๐—”๐—œ ๐—ฎ๐—ด๐—ฒ๐—ป๐˜ ๐—ฐ๐—ฎ๐—ป ๐—ฎ๐—ฐ๐˜ ๐˜„๐—ถ๐˜๐—ต๐—ผ๐˜‚๐˜ ๐—ฎ ๐—ต๐˜‚๐—บ๐—ฎ๐—ป ๐—บ๐—ฎ๐—ธ๐—ถ๐—ป๐—ด ๐˜๐—ต๐—ฒ ๐—ณ๐—ถ๐—ป๐—ฎ๐—น ๐—ฑ๐—ฒ๐—ฐ๐—ถ๐˜€๐—ถ๐—ผ๐—ป, ๐—ฎ๐—ฐ๐—ฐ๐—ผ๐˜‚๐—ป๐˜๐—ฎ๐—ฏ๐—ถ๐—น๐—ถ๐˜๐˜† ๐—ฐ๐—ฎ๐—ป'๐˜ ๐—ฏ๐—ฒ ๐—ฎ๐—ฑ๐—ฑ๐—ฒ๐—ฑ ๐—ฎ๐—ณ๐˜๐—ฒ๐—ฟ ๐˜๐—ต๐—ฒ ๐—ถ๐—ป๐—ฐ๐—ถ๐—ฑ๐—ฒ๐—ป๐˜. ๐—œ๐˜ ๐—ต๐—ฎ๐˜€ ๐˜๐—ผ ๐—ฏ๐—ฒ ๐—ฑ๐—ฒ๐˜€๐—ถ๐—ด๐—ป๐—ฒ๐—ฑ ๐—ถ๐—ป๐˜๐—ผ ๐˜๐—ต๐—ฒ ๐˜€๐˜†๐˜€๐˜๐—ฒ๐—บ. That may be the real architecture challenge of agentic AI. ๐—”๐—ฟ๐—ฒ ๐˜„๐—ฒ ๐—ฏ๐˜‚๐—ถ๐—น๐—ฑ๐—ถ๐—ป๐—ด ๐—ฒ๐—ป๐—ผ๐˜‚๐—ด๐—ต ๐˜๐—ฟ๐—ฎ๐—ฐ๐—ฒ๐—ฎ๐—ฏ๐—ถ๐—น๐—ถ๐˜๐˜† ๐—ถ๐—ป๐˜๐—ผ ๐˜๐—ต๐—ฒ ๐˜€๐˜†๐˜€๐˜๐—ฒ๐—บ๐˜€ ๐˜„๐—ฒ'๐—ฟ๐—ฒ ๐—ด๐—ถ๐˜ƒ๐—ถ๐—ป๐—ด ๐—ฎ๐˜‚๐˜๐—ผ๐—ป๐—ผ๐—บ๐˜† ๐˜๐—ผ? #ArtificialIntelligence #AIAgents #TechnologyLeadership #SoftwareArchitecture
7
232
What if the next lie detector doesnโ€™t ask you a question? It watches you instead. The Pentagon is reportedly exploring an AI-powered polygraph that could use signals such as breathing and heart rate to assess whether someone is lying. I can see the technical appeal. What makes me uneasy is the assumption that those signals can reliably tell us whether someone is telling the truth. AI can identify patterns, ๐—ฏ๐˜‚๐˜ ๐—ฎ ๐—ฝ๐—ฎ๐˜๐˜๐—ฒ๐—ฟ๐—ป ๐—ถ๐˜€๐—ปโ€™๐˜ ๐—ฝ๐—ฟ๐—ผ๐—ผ๐—ณ. If the underlying signals are ambiguous, a more sophisticated model could produce a confident answer without resolving that uncertainty. And once that answer comes with a score and a label, it becomes easier to treat a questionable judgment as an objective finding. Getting something wrong in a chatbot response is one thing. ๐—ช๐—ฟ๐—ผ๐—ป๐—ด๐—น๐˜† ๐—น๐—ฎ๐—ฏ๐—ฒ๐—น๐—น๐—ถ๐—ป๐—ด ๐˜€๐—ผ๐—บ๐—ฒ๐—ผ๐—ป๐—ฒ ๐—ฎ ๐—น๐—ถ๐—ฎ๐—ฟ ๐—ฐ๐—ผ๐˜‚๐—น๐—ฑ ๐—ฎ๐—ณ๐—ณ๐—ฒ๐—ฐ๐˜ ๐˜๐—ต๐—ฒ๐—ถ๐—ฟ ๐—ท๐—ผ๐—ฏ, ๐—ฟ๐—ฒ๐—ฝ๐˜‚๐˜๐—ฎ๐˜๐—ถ๐—ผ๐—ป ๐—ผ๐—ฟ ๐—ณ๐—ฟ๐—ฒ๐—ฒ๐—ฑ๐—ผ๐—บ. Before we give AI that kind of authority, we need evidence that the judgment is reliable and a meaningful way for people to challenge it. ๐—ช๐—ผ๐˜‚๐—น๐—ฑ ๐˜†๐—ผ๐˜‚ ๐˜๐—ฟ๐˜‚๐˜€๐˜ ๐—ฎ๐—ป ๐—”๐—œ ๐—ฎ๐˜€๐˜€๐—ฒ๐˜€๐˜€๐—บ๐—ฒ๐—ป๐˜ ๐—ผ๐—ณ ๐˜„๐—ต๐—ฒ๐˜๐—ต๐—ฒ๐—ฟ ๐˜€๐—ผ๐—บ๐—ฒ๐—ผ๐—ป๐—ฒ ๐—ถ๐˜€ ๐—น๐˜†๐—ถ๐—ป๐—ด? What evidence would you need? #ArtificialIntelligence #ResponsibleAI #AI #TechnologyLeadership #FutureOfTechnology
118
โ€œDonโ€™t click suspicious adsโ€ only gets you so far when displaying one can deliver spyware. Citizen Lab reports that targeted ads were used to infect devices at Al Jazeera with Candiru spyware. It calls the technique AdFection: the ad loads code that can deliver an exploit to a selected target without a click. Discover-Techโ€™s ads reached almost 200,000 websites and apps over four months. Thatโ€™s advertising reach, not evidence that everyone who saw them was infected. How should security advice change when avoiding the click isnโ€™t enough? citizenlab.ca/research/how-dโ€ฆ
112
An AI agent shouldnโ€™t be responsible for enforcing its own limits. Thatโ€™s what caught my attention in @NVIDIAโ€™s Open Agent Safety Platform announcement. OpenShell enforces policies while the agent runs. Sentry adds a separate hardware-based watchdog in the reference design. The model is only part of the system. Context, tools, permissions and execution all shape what an agent can actually do. ๐—ง๐—ต๐—ฒ ๐—ฟ๐—ฒ๐—ฎ๐—น ๐˜€๐˜†๐˜€๐˜๐—ฒ๐—บ ๐—ถ๐˜€ ๐—ฐ๐—น๐—ผ๐˜€๐—ฒ๐—ฟ ๐˜๐—ผ: ๐—บ๐—ผ๐—ฑ๐—ฒ๐—น โ†’ ๐—ฐ๐—ผ๐—ป๐˜๐—ฒ๐˜…๐˜ โ†’ ๐˜๐—ผ๐—ผ๐—น๐˜€ โ†’ ๐—ฝ๐—ฒ๐—ฟ๐—บ๐—ถ๐˜€๐˜€๐—ถ๐—ผ๐—ป๐˜€ โ†’ ๐—ฒ๐˜…๐—ฒ๐—ฐ๐˜‚๐˜๐—ถ๐—ผ๐—ป โ†’ ๐—ณ๐—ฒ๐—ฒ๐—ฑ๐—ฏ๐—ฎ๐—ฐ๐—ธ Once an agent can act on real systems, instructions need to be backed by controls it canโ€™t simply work around. When something goes wrong, โ€œwhy did the model do that?โ€ is only part of the postmortem. Iโ€™d also want to know: โ€ข What authority did we give it? โ€ข Which systems could it reach? โ€ข What actions were allowed? โ€ข What was monitoring them? โ€ข Could we stop it and revoke its access? The model provides capability. The architecture sets the limits on how itโ€™s used. Those limits need to exist ๐—ฏ๐—ฒ๐—ณ๐—ผ๐—ฟ๐—ฒ ๐˜๐—ต๐—ฒ ๐—ฎ๐—ด๐—ฒ๐—ป๐˜ ๐—ด๐—ฒ๐˜๐˜€ ๐—ฎ๐—ฐ๐—ฐ๐—ฒ๐˜€๐˜€ ๐˜๐—ผ ๐—ฟ๐—ฒ๐—ฎ๐—น ๐˜€๐˜†๐˜€๐˜๐—ฒ๐—บ๐˜€. For engineering leaders, this brings familiar systems-design questions into AI: permissions, isolation, monitoring and recovery. Where are you putting independent controls around your agents? Supporting research: arxiv.org/abs/2609.00006 NVIDIAโ€™s approach: nvidianews.nvidia.com/news/oโ€ฆ #AISafety #Cybersecurity #EngineeringLeadership #Nvidia
2
2
159
Oh no! It looks like @PLAUDAI removed a useful feature of automatically labelling the same speakers in every meeting. Now, except for my own voice, I have to go back to labelling each speaker separately for every single meeting! Is this feature removal connected to the voiceprint litigation facing Otter and Fireflies? If privacy concerns drove the decision, explain them. Removing a useful feature and handing the work back to users deserves more than a help-page update. support.plaud.ai/hc/en-us/arโ€ฆ
2
186
Toronto after dark, Niagara in full roar and a few quieter moments in between. A quick trip to Canada with plenty to point the camera at. ๐Ÿ‡จ๐Ÿ‡ฆ #Canada #Travel #TravelPhotography
1
1
163