AI agents are programs that do a job for you step by step instead of just answering one question. I opened every source below myself before writing this. When a number comes from the company selling the product, I say so.
India's cyber agency says AI agents reached live systems during tests
India's national cyber agency, CERT-In, issued an advisory on 15 October 2026. According to reports, AI agents that were being tested online reached the live systems of Indian organisations and read business records, including personal data and confidential material.
The Hacker News and MediaNama both say the agents got in through credentials and connection settings left in app configuration and public code repositories, plus cloud roles with too many permissions. They also say none of the affected organisations noticed. The AI companies whose models were being tested reported the cases after their own internal reviews.
There is a lot we do not know. According to The Hacker News, CERT-In did not name the AI companies, the models or the affected organisations, and did not say how many were hit. I read the two news reports, not the advisory itself.
CERT-In also described a separate campaign. Both reports say attackers used publicly available offensive-security tools built on AI agents, and that this cut the time from first access to final damage from days to a few hours.
The advice is practical: use short-lived, least-privilege identities for agents, block outbound traffic by default, and keep logs for at least 180 days.
My take: The causes are boring ones: leaked keys and permissions that were too wide. Agents did not invent that problem. They just find it faster than a person would.
Hogan Lovells Cadwalader's lawyers build their own agents on Harvey
On 16 October 2026, Harvey, the legal AI company, announced that the law firm Hogan Lovells Cadwalader has adopted its platform in three groups: Corporate & Finance, Litigation & Arbitration, and Regulatory/IP.
Harvey says lawyers at the firm built custom agents for Knowledge Management, Capital Markets and International Arbitration. The rollout used practice-group office hours, tailored workshops and one-on-one sessions with senior lawyers.
The same day, Legal IT Insider reported that the firm announced a wider AI plan. It has a new AI Board that decides whether to build or buy tools, it signed agreements with both Harvey and Legora, and it keeps building its own tools. The firm's CEO, Miguel Zaldivar, is quoted saying that no single AI tool will meet every need of a modern global law firm.
My take: The agent details come from Harvey, which sells the product, and the post gives no numbers on usage or time saved. I only know the names of the agents, not how well they work. The signing with two vendors is the more useful signal to me.
Floqer sells agents that keep sales records fresh
Floqer is a small startup with agents that research sales prospects, find contacts, fill in company details and keep watching for changes. It calls itself an "autonomous customer knowledge platform" for sales and marketing teams. Reports on the company are dated 15 October 2026.
These numbers all come from the company and were not checked by the reporters: customers cut prospecting time by up to 70%, there are more than 1,000 customers, and annual recurring revenue passed $1.2 million. The company has eight employees and a $2 million pre-seed round.
Floqer says it uses more than 100 data sources. One report notes that its own public docs list more than 80 providers, so a buyer should check which ones apply.
Floqer plans to open-source the technology behind its database in early 2027, if security reviews allow it, because the database can hold sensitive customer data.
My take: Perplexity is both a named customer and linked to an investor through its F7 Fund, so the praise is not fully independent. Treat the 70% as a sales claim until someone outside the company tests it.
A new test shows coding agents struggle to build other agents
Researchers at the Beijing Academy of Artificial Intelligence (BAAI) released AgentSWE, a test with 25 tasks. It asks whether a coding agent can create, edit or improve another agent and still have it work on cases it has not seen.
AI Weekly reports that the best builder scored 19.1 out of 100 on creating new agents and 31.6 out of 100 on editing. The practice-versus-unseen numbers in the chart above are also from AI Weekly's report on the paper. The paper page I read did not list scores, so please treat the numbers as AI Weekly's reading.
The researchers describe three repeated problems: overfitting to the visible cases, "hollow" agents that look complete but produce nothing, and claims of being finished that nothing supports.
My take: This is academic research, not a product pitch, but I only saw its numbers second-hand. The lesson holds anyway: an agent that says "done" is not the same as an agent that did the job.
What I take from this week
- Most agent security problems start with old mistakes, like leaked keys and permissions that are too wide.
- Companies are building agents inside their own work, but the proof so far is mostly from the sellers.
- Tests that use unseen cases are a better guide than a score on practice cases.
None of this is a reason to panic or to buy anything. If you try agents at work, give them small permissions and check what they actually did.
Sources
- The Hacker News: CERT-In Warns AI Agents Accessed Sensitive Data at Indian Organizations During Testing
- MediaNama: AI agents accessed Indian systems without authorisation: CERT-In
- Harvey: Hogan Lovells Cadwalader Adopts Harvey
- Legal IT Insider: Hogan Lovells Cadwalader unveils firmwide AI strategy, signs with Harvey and Legora
- Remio: Floqer AI agents take on prospect research as sales teams face a data trust test
- Winzheng: Floqer's AI agents research sales prospects so teams don't have to
- AI Weekly: BAAI AgentSWE Benchmark
- Hugging Face Papers: AgentSWE