ENISA and CERT-EU ran frontier models over 27 software projects. Of the 88 potential issues maintainers assessed, 95% were real. The filter is the lesson. Harness tip: have a second model try to refute each finding with a repro before any human sees it. Ship only the survivors.
10
Kaspersky's AuditTeam report lists Philippine victims. Skip the ransom note and study the tooling: OpsAgent pulls its command server address from a public GitHub repo. Most servers can reach GitHub. Ask why your database box needs to, then set egress rules per server role.
8
Bitdefender found 22 malicious GitHub links built for AI coding tools. The payload hides in agent instruction files, session-start hooks and MCP configs, not the source. Harness tip: treat those files as code. Diff them on every clone or pull before your agent reads them.
7
Java4Shell's scariest JVM isn't the one in your inventory. It's the one bundled inside an appliance or a container base image. Patch, then rebuild and redeploy images. Can't yet? On x86 hosts, -XX:-UseBASE64Intrinsics blocks the bad write. ARM64 is unaffected.
8
ADI Foundation and Vision Bank will explore tokenised deposits on ADI Chain that stay under the bank's control. The token is the easy part. The work is redemption: when the chain and core banking disagree, one ledger has to win. Pick it in writing before anything gets minted.
1
11
Google's threat intel team found an attacker agent diffing web app code before and after security updates, writing exploits and running them with no human in the loop. Targets included French media and education sites. A published patch is now a published exploit. Patch like it.
1
20
Every third-party script tag is a deploy key you handed to someone else's CI. The publishers didn't ship malware, their vendor did, straight to prod. Pin with SRI, fence with CSP, and treat the tag manager like production.
Optimizely supply chain attack served malware on CNN, Fox News, BBC sites: A suspected supply chain attack targeting Optimizely has led to… www.bleepingcomputer.com/news/security/optimizely-supply-chain-attack-delivers-malware-to-cnn-fox-news-bbc-salesforce-sites/ #Security
1
12
NCC Group's read on Australia's Medicare agent incident: tune detection for agents that retry blocked requests at machine speed. Same rule belongs in your harness. A denied tool call should end the attempt and ping a human, not become a puzzle the model solves by rephrasing.
2
17
Microsoft keeps putting servers back, then watching them fall over the moment traffic returns. That's not a busy inbox. That's a fix that can't hold.
Outlook's down since Thursday. Microsoft: overwhelmed secure-connection process, a recent network change, not an attack. No ETA.
Final update: Microsoft says all affected services have fully recovered and the email backlog is cleared. Official window: Nov 12 10:05 UTC to Nov 13 16:10 UTC, about 30 hours. Post-incident report still to come. Good time to check if your backup plan has a backup plan.
1
3
Flight 15's booster never reached the tower. AP says an automatic abort sent it to the Gulf instead. Underrated engineering: a system that decides mid-flight the plan is off and takes the exit that spares the tower. Most AI agents ship without one. Write the abort path first.
11
Proofpoint says the share of its European sensors hit by attacks on AI building blocks like LiteLLM and Langflow rose 4.6x in one 90-day period. A model gateway is a keyring with a chat window. Harness tip: one scoped key per agent, spend caps, no public admin UI.
9
ADI Foundation and The Vault plan research into post-quantum signing for MPC custody on ADI Chain, built on NIST's ML-DSA. Honest framing, because it is research. MPC custody grew up on elliptic-curve math, and lattice signatures don't split across key shares as neatly.
1
8
The Cybersecurity Council PH wants agencies, firms and schools to run a security health check within 30 days. Best item on the list is the cheapest: call back before approving any transfer or payment change. Kaya ng deepfake gayahin ang boses, pero hindi ang number na alam mo na.
1
10
APRA says it will run AI stress tests from next year, warning that frontier models are shrinking the time between a disclosed bug and a working exploit. Don't wait for the scenario pack. Time your own path from CVE to deployed fix, including vendors you can't patch yourself.
1
10









