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 serverโs 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 canโt 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. ๐ข๐ป๐ฐ๐ฒ ๐๐ต๐ฒ๐ป ๐๐ต๐ฒ ๐๐ผ๐ฟ๐ธ ๐ฟ๐๐ป๐ ๐ฎ๐ป๐ฑ ๐ฎ๐ด๐ฎ๐ถ๐ป ๐๐ต๐ฒ๐ป ๐๐ผ๐ ๐ฑ๐ฒ๐น๐ฒ๐๐ฒ ๐ถ๐. Thatโs what caught my attention in CircleCIโs 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. ๐ง๐ต๐ถ๐ ๐ถ๐ป๐ฐ๐ฟ๐ฒ๐ฎ๐๐ฒ๐ฑ ๐๐ต๐ฒ ๐๐ผ๐ฟ๐ธ ๐ฃ๐ผ๐๐๐ด๐ฟ๐ฒ๐ฆ๐ค๐ ๐ต๐ฎ๐ฑ ๐๐ผ ๐ฑ๐ผ, ๐ฑ๐ฟ๐ถ๐๐ถ๐ป๐ด ๐๐ฃ๐จ ๐๐๐ฎ๐ด๐ฒ ๐ต๐ถ๐ด๐ต๐ฒ๐ฟ. 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 ๐ฟ๐ฒ๐๐ฒ๐ป๐๐ถ๐ผ๐ป ๐ฝ๐ผ๐น๐ถ๐ฐ๐ถ๐ฒ๐ ๐ฎ๐ฟ๐ฒ ๐ฎ๐น๐๐ผ ๐ฐ๐ฎ๐ฝ๐ฎ๐ฐ๐ถ๐๐ ๐ฑ๐ฒ๐ฐ๐ถ๐๐ถ๐ผ๐ป๐. Choosing how long to keep data determines when your infrastructure has to do the work of removing it. In a MarTech platform, ๐๐ฒ๐๐๐ฒ๐ฟ๐ฑ๐ฎ๐โ๐ ๐ฟ๐ฒ๐๐ฒ๐ป๐๐ถ๐ผ๐ป ๐ฑ๐ฒ๐น๐ฒ๐๐ฒ๐ ๐ฐ๐ผ๐๐น๐ฑ ๐ฐ๐ผ๐บ๐ฝ๐ฒ๐๐ฒ ๐๐ถ๐๐ต ๐๐ผ๐ฑ๐ฎ๐โ๐ ๐ฎ๐๐ฑ๐ถ๐ฒ๐ป๐ฐ๐ฒ ๐ฟ๐ฒ๐ณ๐ฟ๐ฒ๐๐ต๐ฒ๐ ๐ฎ๐ป๐ฑ ๐น๐ถ๐๐ฒ ๐ฐ๐๐๐๐ผ๐บ๐ฒ๐ฟ ๐พ๐๐ฒ๐ฟ๐ถ๐ฒ๐.
When evaluating a platformโs performance, Iโd 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 ๐ช๐๐ archiving before customers notice delays?
CircleCI eventually paused incoming and retry work to let the database recover, then gradually increased the load. ๐๐ฎ๐ฐ๐ธ๐ด๐ฟ๐ผ๐๐ป๐ฑ ๐๐ผ๐ฟ๐ธ ๐ป๐ฒ๐ฒ๐ฑ๐ ๐ฎ ๐ฝ๐ฟ๐ผ๐ฑ๐๐ฐ๐๐ถ๐ผ๐ป ๐ฏ๐๐ฑ๐ด๐ฒ๐ ๐๐ผ๐ผ.
Which cleanup job in your system could disrupt live traffic?
[CircleCIโs postmortem](discuss.circleci.com/t/post-โฆ)
\#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. Thereโs 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 weโve paid BESCOM for the repair, paid an electrician for repeated visits and now have damaged appliances to deal with. Thatโs on top of the time and patience spent getting a problem fixed that we didnโt 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/flippeโฆ
#FlipperZero #Insta360
1
1
194
Every year, I find myself reminding my team of one thing: ๐ญ๐ก๐๐ซ๐'๐ฌ ๐ฉ๐ฅ๐๐ง๐ญ๐ฒ ๐จ๐ ๐ญ๐ข๐ฆ๐ ๐ญ๐จ ๐ ๐๐ญ ๐๐๐๐ค ๐ญ๐จ ๐๐ฎ๐ข๐ฅ๐๐ข๐ง๐ . 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 ๐ฌ๐ญ๐๐ฉ ๐๐ฐ๐๐ฒ, ๐ฌ๐ฅ๐จ๐ฐ ๐๐จ๐ฐ๐ง ๐๐ง๐ ๐๐ง๐ฃ๐จ๐ฒ ๐ญ๐ก๐ ๐ฆ๐จ๐ฆ๐๐ง๐ญ 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! ๐ช
P.S. The work can wait. Make the most of the celebrations!
2
91
What if we could ๐ฎ๐ง๐๐๐ซ๐ฌ๐ญ๐๐ง๐ ๐ฅ๐๐ ๐๐๐ฒ ๐ฌ๐จ๐๐ญ๐ฐ๐๐ซ๐ 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:
โข Recover business rules hidden in legacy code.
โข Trace dependencies to understand how different components interact.
โข 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 ๐ฆ๐จ๐ซ๐ ๐ฉ๐ซ๐๐๐ข๐ฌ๐, ๐ญ๐๐ซ๐ ๐๐ญ๐๐ ๐๐ง๐ ๐ฏ๐๐ซ๐ข๐๐ข๐๐๐ฅ๐.
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-enginโฆ
#AI #rea #SoftwareModernisation #ReverseEngineering #SoftwareEngineering
2
98
Vinu Thomas retweeted
โผ๏ธWARNING - 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/opโฆ
A critical vulnerability lands. What needs fixing? Cloudflareโs Zero Wire offers short briefings on whoโs affected, whatโs confirmed and what to do next, with links to the evidence.
radar.cloudflare.com/securitโฆ
62
Two iconic bridges, one timeless river. Evenings by the Hooghly never get old.
#KolkataDiaries #VidyasagarSetu #CityOfJoy #prinsepghat #travel
81
๐ข๐ป๐ฒ ๐๐ต๐ถ๐ป๐ด ๐๐ต๐ฎ๐ ๐ฐ๐ฎ๐ป ๐ฐ๐ฎ๐๐ฐ๐ต ๐๐ ๐ผ๐ณ๐ณ ๐ด๐๐ฎ๐ฟ๐ฑ ๐ถ๐ป ๐๐ฒ๐ฐ๐ต๐ป๐ผ๐น๐ผ๐ด๐ ๐ฝ๐ฟ๐ผ๐ท๐ฒ๐ฐ๐๐ ๐ถ๐ ๐ต๐ผ๐ ๐บ๐๐ฐ๐ต ๐๐ผ๐ฟ๐ธ ๐ฟ๐ฒ๐บ๐ฎ๐ถ๐ป๐ ๐ฎ๐ณ๐๐ฒ๐ฟ ๐๐ต๐ฒ ๐ณ๐ฒ๐ฎ๐๐๐ฟ๐ฒ ๐ถ๐๐๐ฒ๐น๐ณ ๐ถ๐ ๐ฏ๐๐ถ๐น๐.
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 ๐๐ผ๐ฟ๐ธ๐ถ๐ป๐ด ๐ณ๐ฒ๐ฎ๐๐๐ฟ๐ฒ ๐ถ๐ ๐ป๐ผ๐ ๐ฒ๐ป๐ผ๐๐ด๐ต ๐๐ผ ๐ท๐๐๐๐ถ๐ณ๐ ๐๐ต๐ฒ ๐ถ๐ป๐๐ฒ๐๐๐บ๐ฒ๐ป๐. The solution needs to deliver measurable value at a sustainable cost.
Before committing to a wider rollout, Iโd focus on three questions:
๐ง๐ฒ๐ฐ๐ต๐ป๐ถ๐ฐ๐ฎ๐น ๐ฑ๐ฒ๐ฝ๐ฒ๐ป๐ฑ๐ฒ๐ป๐ฐ๐ถ๐ฒ๐: Are the integrations, data quality and access controls reliable enough for production?
๐๐ผ๐๐ ๐ฝ๐ฒ๐ฟ ๐ผ๐๐๐ฐ๐ผ๐บ๐ฒ: What does each successfully resolved request cost, and how does that change as usage grows?
๐๐๐ถ๐ฑ๐ฒ๐ป๐ฐ๐ฒ ๐ณ๐ผ๐ฟ ๐๐ฐ๐ฎ๐น๐ฒ: 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.
๐ง๐ต๐ฒ ๐ฟ๐ฒ๐ฎ๐น ๐๐ฒ๐๐ ๐ถ๐๐ป'๐ ๐๐ต๐ฒ๐๐ต๐ฒ๐ฟ ๐๐ฒ ๐ฐ๐ฎ๐ป ๐ฏ๐๐ถ๐น๐ฑ ๐๐ต๐ฒ ๐๐ฒ๐ฐ๐ต๐ป๐ผ๐น๐ผ๐ด๐. ๐๐'๐ ๐๐ต๐ฒ๐๐ต๐ฒ๐ฟ ๐๐ต๐ฒ ๐ฟ๐ฒ๐๐๐น๐๐ ๐ท๐๐๐๐ถ๐ณ๐ ๐๐ฎ๐ธ๐ถ๐ป๐ด ๐ถ๐ ๐ณ๐๐ฟ๐๐ต๐ฒ๐ฟ.
How do you decide when a promising technology project is ready to scale?
#AIImplementation #TechnologyStrategy #EnterpriseAI #ITLeadership
1
3
177
๐ ๐ณ๐ฎ๐๐๐ฒ๐ฟ ๐๐ฃ๐จ ๐ฑ๐ผ๐ฒ๐๐ป'๐ ๐ฎ๐น๐๐ฎ๐๐ ๐บ๐ฒ๐ฎ๐ป ๐ฎ ๐ฐ๐ต๐ฒ๐ฎ๐ฝ๐ฒ๐ฟ ๐ฑ๐ฎ๐๐ฎ ๐ฐ๐ฒ๐ป๐๐ฟ๐ฒ.
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 ๐๐ผ๐ฟ๐ฒ ๐๐ถ๐ฟ๐๐ ๐ฎ๐ฝ๐ฝ๐ฟ๐ผ๐ฎ๐ฐ๐ต. 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. ๐ง๐ต๐ฒ ๐ถ๐ป๐๐๐ฟ๐๐ฐ๐๐ถ๐ผ๐ป-๐๐ฒ๐ ๐ฎ๐ฟ๐ฐ๐ต๐ถ๐๐ฒ๐ฐ๐๐๐ฟ๐ฒ ๐ฎ๐ณ๐ณ๐ฒ๐ฐ๐๐ ๐๐ต๐ถ๐ฐ๐ต ๐๐ผ๐ณ๐๐๐ฎ๐ฟ๐ฒ ๐ฎ ๐ฝ๐ฟ๐ผ๐ฐ๐ฒ๐๐๐ผ๐ฟ ๐ฐ๐ฎ๐ป ๐ฟ๐๐ป. 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 ๐ฃ๐ฒ๐ฟ๐ณ๐ผ๐ฟ๐บ๐ฎ๐ป๐ฐ๐ฒ: Do the gains hold up under the workloads we actually run?
โ ๐ฃ๐ผ๐๐ฒ๐ฟ: How much useful work are we getting for the electricity consumed?
โ ๐๐ฑ๐ผ๐ฝ๐๐ถ๐ผ๐ป: 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. ๐ง๐ต๐ฒ ๐๐ผ๐ณ๐๐๐ฎ๐ฟ๐ฒ ๐ฒ๐ฐ๐ผ๐๐๐๐๐ฒ๐บ ๐ฎ๐ป๐ฑ ๐๐ต๐ฒ ๐ฐ๐ผ๐๐ ๐ผ๐ณ ๐ฏ๐ฟ๐ถ๐ป๐ด๐ถ๐ป๐ด ๐ถ๐ ๐ถ๐ป๐๐ผ ๐ฝ๐ฟ๐ผ๐ฑ๐๐ฐ๐๐ถ๐ผ๐ป ๐บ๐ฎ๐๐๐ฒ๐ฟ ๐๐ผ๐ผ.
@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/transactioโฆ
#Semiconductors #DataCenters #CPUDesign #DataCenterEfficiency #ProcessorDesign
101
Vinu Thomas retweeted
This is Genie 4, our most realistic and interactive world model yet. ๐
Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.
When the sky wakes up in flames and the geese welcome in a brand-new day.
Morning magic over Lake Simcoe, Barrie. ๐จ๐ฆ #MorningGlow #SkyOnFire #travelmoments #barrieontario
๐ช๐ฒโ๐๐ฒ ๐๐ฝ๐ฒ๐ป๐ ๐ฑ๐ฒ๐ฐ๐ฎ๐ฑ๐ฒ๐ ๐ฑ๐ฒ๐๐ถ๐ด๐ป๐ถ๐ป๐ด ๐๐๐๐๐ฒ๐บ๐ ๐ฎ๐ฟ๐ผ๐๐ป๐ฑ ๐ฎ ๐๐ถ๐บ๐ฝ๐น๐ฒ ๐ฎ๐๐๐๐บ๐ฝ๐๐ถ๐ผ๐ป: ๐ถ๐ณ ๐๐ผ๐บ๐ฒ๐๐ต๐ถ๐ป๐ด ๐ด๐ผ๐ฒ๐ ๐๐ฟ๐ผ๐ป๐ด, ๐๐ฒ ๐ฐ๐ฎ๐ป ๐๐ฟ๐ฎ๐ฐ๐ฒ ๐๐ต๐ฒ ๐ฑ๐ฒ๐ฐ๐ถ๐๐ถ๐ผ๐ป.
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.
Weโre already seeing the cracks:
โข AI can confidently make the wrong call while users treat it as authoritative.
โข Agents can take actions their creators or users never explicitly intended.
The obvious question is: ๐๐ต๐ผ ๐ถ๐ ๐น๐ถ๐ฎ๐ฏ๐น๐ฒ?
The more important question for technology leaders is: ๐ฐ๐ฎ๐ป ๐ผ๐๐ฟ ๐๐๐๐๐ฒ๐บ๐ ๐ฒ๐๐ฒ๐ป ๐ฒ๐
๐ฝ๐น๐ฎ๐ถ๐ป ๐ต๐ผ๐ ๐๐ต๐ฒ ๐ฑ๐ฒ๐ฐ๐ถ๐๐ถ๐ผ๐ป ๐ต๐ฎ๐ฝ๐ฝ๐ฒ๐ป๐ฒ๐ฑ?
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.
๐๐ณ ๐ฎ๐ป ๐๐ ๐ฎ๐ด๐ฒ๐ป๐ ๐ฐ๐ฎ๐ป ๐ฎ๐ฐ๐ ๐๐ถ๐๐ต๐ผ๐๐ ๐ฎ ๐ต๐๐บ๐ฎ๐ป ๐บ๐ฎ๐ธ๐ถ๐ป๐ด ๐๐ต๐ฒ ๐ณ๐ถ๐ป๐ฎ๐น ๐ฑ๐ฒ๐ฐ๐ถ๐๐ถ๐ผ๐ป, ๐ฎ๐ฐ๐ฐ๐ผ๐๐ป๐๐ฎ๐ฏ๐ถ๐น๐ถ๐๐ ๐ฐ๐ฎ๐ป'๐ ๐ฏ๐ฒ ๐ฎ๐ฑ๐ฑ๐ฒ๐ฑ ๐ฎ๐ณ๐๐ฒ๐ฟ ๐๐ต๐ฒ ๐ถ๐ป๐ฐ๐ถ๐ฑ๐ฒ๐ป๐. ๐๐ ๐ต๐ฎ๐ ๐๐ผ ๐ฏ๐ฒ ๐ฑ๐ฒ๐๐ถ๐ด๐ป๐ฒ๐ฑ ๐ถ๐ป๐๐ผ ๐๐ต๐ฒ ๐๐๐๐๐ฒ๐บ.
That may be the real architecture challenge of agentic AI.
๐๐ฟ๐ฒ ๐๐ฒ ๐ฏ๐๐ถ๐น๐ฑ๐ถ๐ป๐ด ๐ฒ๐ป๐ผ๐๐ด๐ต ๐๐ฟ๐ฎ๐ฐ๐ฒ๐ฎ๐ฏ๐ถ๐น๐ถ๐๐ ๐ถ๐ป๐๐ผ ๐๐ต๐ฒ ๐๐๐๐๐ฒ๐บ๐ ๐๐ฒ'๐ฟ๐ฒ ๐ด๐ถ๐๐ถ๐ป๐ด ๐ฎ๐๐๐ผ๐ป๐ผ๐บ๐ ๐๐ผ?
#ArtificialIntelligence #AIAgents #TechnologyLeadership #SoftwareArchitecture
7
232
A cheaper cloud bill can come with an expensive exit. Integrations and team skills tie you to a platform. Before you commit, ask what leaving would take.
vinuthomas.com/how-do-technoโฆ
1
120
What if the next lie detector doesnโt 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, ๐ฏ๐๐ ๐ฎ ๐ฝ๐ฎ๐๐๐ฒ๐ฟ๐ป ๐ถ๐๐ปโ๐ ๐ฝ๐ฟ๐ผ๐ผ๐ณ.
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.
๐ช๐ฟ๐ผ๐ป๐ด๐น๐ ๐น๐ฎ๐ฏ๐ฒ๐น๐น๐ถ๐ป๐ด ๐๐ผ๐บ๐ฒ๐ผ๐ป๐ฒ ๐ฎ ๐น๐ถ๐ฎ๐ฟ ๐ฐ๐ผ๐๐น๐ฑ ๐ฎ๐ณ๐ณ๐ฒ๐ฐ๐ ๐๐ต๐ฒ๐ถ๐ฟ ๐ท๐ผ๐ฏ, ๐ฟ๐ฒ๐ฝ๐๐๐ฎ๐๐ถ๐ผ๐ป ๐ผ๐ฟ ๐ณ๐ฟ๐ฒ๐ฒ๐ฑ๐ผ๐บ.
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.
๐ช๐ผ๐๐น๐ฑ ๐๐ผ๐ ๐๐ฟ๐๐๐ ๐ฎ๐ป ๐๐ ๐ฎ๐๐๐ฒ๐๐๐บ๐ฒ๐ป๐ ๐ผ๐ณ ๐๐ต๐ฒ๐๐ต๐ฒ๐ฟ ๐๐ผ๐บ๐ฒ๐ผ๐ป๐ฒ ๐ถ๐ ๐น๐๐ถ๐ป๐ด?
What evidence would you need?
#ArtificialIntelligence #ResponsibleAI #AI #TechnologyLeadership #FutureOfTechnology
118
โDonโt 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-Techโs ads reached almost 200,000 websites and apps over four months. Thatโs advertising reach, not evidence that everyone who saw them was infected.
How should security advice change when avoiding the click isnโt enough?
citizenlab.ca/research/how-dโฆ
112
An AI agent shouldnโt be responsible for enforcing its own limits.
Thatโs what caught my attention in @NVIDIAโs 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.
๐ง๐ต๐ฒ ๐ฟ๐ฒ๐ฎ๐น ๐๐๐๐๐ฒ๐บ ๐ถ๐ ๐ฐ๐น๐ผ๐๐ฒ๐ฟ ๐๐ผ:
๐บ๐ผ๐ฑ๐ฒ๐น โ ๐ฐ๐ผ๐ป๐๐ฒ๐
๐ โ ๐๐ผ๐ผ๐น๐ โ ๐ฝ๐ฒ๐ฟ๐บ๐ถ๐๐๐ถ๐ผ๐ป๐ โ ๐ฒ๐
๐ฒ๐ฐ๐๐๐ถ๐ผ๐ป โ ๐ณ๐ฒ๐ฒ๐ฑ๐ฏ๐ฎ๐ฐ๐ธ
Once an agent can act on real systems, instructions need to be backed by controls it canโt simply work around.
When something goes wrong, โwhy did the model do that?โ is only part of the postmortem.
Iโd also want to know:
โข What authority did we give it?
โข Which systems could it reach?
โข What actions were allowed?
โข What was monitoring them?
โข Could we stop it and revoke its access?
The model provides capability.
The architecture sets the limits on how itโs used.
Those limits need to exist ๐ฏ๐ฒ๐ณ๐ผ๐ฟ๐ฒ ๐๐ต๐ฒ ๐ฎ๐ด๐ฒ๐ป๐ ๐ด๐ฒ๐๐ ๐ฎ๐ฐ๐ฐ๐ฒ๐๐ ๐๐ผ ๐ฟ๐ฒ๐ฎ๐น ๐๐๐๐๐ฒ๐บ๐.
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
NVIDIAโs approach:
nvidianews.nvidia.com/news/oโฆ
#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/arโฆ
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. ๐จ๐ฆ
#Canada #Travel #TravelPhotography
1
1
163















