Hey friends 👋
Two things happened this week that made me sit up: security teams just got an army of AI coworkers, and Meta might start selling the one thing everyone assumed it could never have enough of. Let's dig in.
Cognition built an AI team for your security backlog, and the numbers are wild
Here's a problem most people outside security don't think about: as AI writes more and more of the world's code, someone still has to check that code for holes. And that job is getting brutal. Reports suggest flagged security issues at some companies have gone from around 1,000 a month to over 10,000 in half a year, and a huge chunk of that is just noise, scanners crying wolf because they can't tell a real threat from a harmless pattern match.
Cognition, the team behind the coding agent Devin, thinks it has an answer: don't send one scanner, send a swarm. Devin Security Swarm splits a codebase into pieces and sends a small army of agents after it in parallel. Each one investigates its slice, and instead of just flagging "this looks sketchy," it actually tries to exploit the bug in a sandboxed copy of the code to prove it's real. If it's real, Devin writes the fix and opens a pull request itself.
The benchmark results are the part that got my attention. Cognition ran it against 50 known vulnerabilities, real bugs from public GitHub security advisories across 14 languages, and the swarm found 36 of them. That's more than any other AI scanner tested, at about 30% less cost per finding. Three vulnerabilities, including a nasty sandbox bypass and a deserialization flaw in Spring Kafka, were caught by nobody else.
If this holds up outside of a benchmark, it's a meaningful shift: security teams stop drowning in maybes and start getting a short list of confirmed, exploitable problems with the fix already written.
Meta might rent out its AI compute, and Wall Street is thrilled about it
For years, Meta had one job for its mountain of AI infrastructure: make Instagram and Facebook smarter, stickier, more profitable. It never sold that infrastructure to anyone else. That's the neocloud crowd's job, companies like CoreWeave and Nebius that exist specifically to rent out GPUs to whoever needs them, including Meta itself.
That might be about to change. Bloomberg is reporting, based on people familiar with internal talks, that Meta is building a cloud business internally called Meta Compute. The pitch is simple: Meta has been spending an eye-watering amount on AI infrastructure (2026 guidance is $125 to $145 billion, roughly double last year), and if some of that capacity sits idle, why not sell it? Two options are reportedly on the table, letting outside developers run on Meta's models the way Amazon's Bedrock works, or just renting out raw GPU horsepower.
Nothing here is confirmed. Meta hasn't said a word publicly, and Bloomberg's own reporting notes the plan could still change. But the market didn't wait for confirmation: Meta's stock jumped more than 10% in a single day on the news, adding over $100 billion in value, while CoreWeave and Nebius both dropped sharply. Which makes sense: if Meta starts selling its own leftover compute, two of the biggest customers of the neocloud business just became its competitors.
Zuckerberg actually telegraphed this back in May, saying selling excess capacity was "definitely on the table" if Meta ended up overbuilding. Looks like the "if" quietly turned into a "maybe, soon."
A few more things worth knowing
Cognition calls its approach Agentic MapReduce: one agent maps out the whole codebase first, then hands out the work in chunks so the parallel agents aren't stepping on each other. It's the same divide-and-conquer idea from distributed computing, just applied to hunting bugs.
Meta isn't a small player in the compute-renting world already, just on the buying side. It's reportedly committed something like $48 billion to renting GPUs from CoreWeave and Nebius combined, which is what makes a pivot to competing with them such a notable reversal.
The people reportedly steering Meta Compute: infrastructure chief Santosh Janardhan, Meta Superintelligence Labs lead Daniel Gross, and president Dina Powell McCormick.
That's what stood out to me today. Reply and tell me what caught your eye, I read everything.
Talk tomorrow,
Hatman 🎩