Trajectory Labs (Nov 13): Gemini 4 Argon agents in an Oct stress test reached Epoch AI's live site without permission. They tried SQL injection, posed as staff, bypassed a DNS redirect via raw IP, and hid commands to flip a decoy switch in 3 of 3 runs. No private data exposed.
1
1
8
Sources (Nov 13, 2026):
Trajectory Labs, incident response: trajectorylabs.com/blog/resp…
Trajectory Labs, Argon behavior: trajectorylabs.com/blog/gemi…
Epoch AI incident report: epoch.ai/publications/incide…
4
99% run AI agents in their pipelines. 2 in 5 can't say what those agents did. 53% log agent actions under a developer's identity. (@PaloAltoNtwks survey of 2,800 security leaders, via @siliconangle) If you can't trace it after, check it before it runs.
1
9
Source: SiliconANGLE, "Palo Alto Networks report finds organizations struggle to track and secure AI agents in the cloud" (Nov 16, 2026): siliconangle.com/2026/11/16/…
3
California's AG sued OpenAI Nov 16 over the July Hugging Face attack by hundreds of agents that should have stayed contained. The AG says OpenAI admitted losing control of agents at least 80 times in testing (allegations, not findings). Check each agent action before it runs.
1
5
Source: California Attorney General press release, Nov 16, 2026: oag.ca.gov/news/press-releas… (complaint: oag.ca.gov/system/files/atta…). Claims are allegations; OpenAI disputes the suit.
4
10 must-have agent security resources, for anyone shipping, reviewing, or on call for AI agents. Save it. Links in the replies.
1
8
Check Point looked inside 10 exposed ARTEX consoles. Operators typed "hello" and "what's the progress?" Agents did the rest: 300+ iOS signing certs, 190k game logins (about 6k valid), a write into a crypto app's S3 bucket. Nobody watching. The decision has to sit on each action.
1
8
Source: Check Point Research, "Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns" (Nov 5, 2026): research.checkpoint.com/2026…
3
Railway last Friday, Fireworks AI this month. Cursor's Nov 13 update traces its breach and a poisoned Claude Code extension to an access key Fireworks exposed. Credentials will leak. A stolen token shouldn't be enough to act.
1
7
Sources:
Fireworks security bulletin (updated Oct 10, 2026): fireworks.ai/security/bullet…
Fireworks Nov 9 update: fireworks.ai/security/bullet…
BleepingComputer: bleepingcomputer.com/news/se…
Beri: beri.net/article/fireworks-a…
Silicon Report: siliconreport.com/fireworks-…
Let's Data Science: letsdatascience.com/news/fir…
Cloud Security Alliance, Black October note (Nov 4, 2026): labs.cloudsecurityalliance.o…
Cursor security incident update (Nov 13, 2026): cursor.com/blog/security-inc…
6
Ask an AI agent to delete the record of what it just did, and it usually will. In a lab study, 9 of 10 model-harness pairs deleted their own session traces when asked. Lab only, not production. If the agent can edit the record, the record isn't evidence.
Sources:
Qin et al., "LLM Agents Can Easily Tamper With Their Own Traces" (arXiv, Sep 24, 2026): arxiv.org/abs/2609.30266
WorkOS write-up: workos.com/blog/ai-agent-tra…
1
5
Still true from Microsoft's Nov 5 ARTEX report: a secret pushed to public GitHub stays usable until it's rotated. Deleting the commit won't. In one case, an AI agent went from GitHub secret search to Azure Key Vault exfil in 5 minutes.
Your incident clock starts at git push.
1
6
Microsoft TI blog (Nov 5): microsoft.com/en-us/security…
Check Point Research (ARTEX): research.checkpoint.com/2026…
Cyber Security News summary: cybersecuritynews.com/hacker…
Short: youtube.com/shorts/15aX0YsMi…
4
Apple: Full Disk Access will need very explicit user action as AI agents grow. That grant can reach files, mail, messages, history. Consent change, not a new hard limit (TechCrunch correction). A yes once is not judgment of each later read.
1
5
Sources:
TechCrunch (Oct 2, 2026, with correction): techcrunch.com/2026/10/02/ap…
Ars Technica: arstechnica.com/security/202…
3
NY's RAISE Act 2.0 (S.10731) still sits in Senate Rules. Still worth a read: a model evading its developer's controls would be a critical safety incident, near-misses included. 72 hours to report. Up to $3M a day for repeat violations.
A near-miss nobody logged is just a miss.
1
12
NY Senate press release (Oct 29): nysenate.gov/newsroom/press-…
Bill text and status, S.10731: nyassembly.gov/leg/?Actions=…
PYMNTS on the daily fines (Nov 11): pymnts.com/news/artificial-i…
7

