Google and Amazon report AI agents targeting security systems

Google’s Threat Intelligence Group says it found exploit_agent on an exposed staging server, describing it as the first case it has observed of threat actors fully automating vulnerability research and exploitation. The framework compares code before and after a security patch, uses large language models to identify flaws, and attempts to exploit them without human input. Google also reported other AI-supported activity, including open-source frameworks CAIRN and PentAGI used in attacks, and a platform being developed with the Gemini API to manage more than 100,000 social media accounts. Google says it disabled accounts and assets linked to misuse of its services and strengthened safeguards.

Amazon’s threat intelligence team found records of AI agents that it believes, with moderate confidence, ran on the ARTEX framework. The agents searched for exposed AI credentials, checked cloud identities and permissions, and shared information through a persistent knowledge graph. One campaign targeted service keys and applications including LiteLLM, Dify, FastGPT and Open WebUI; another used real-time deepfake video and forged ID images to test identity checks at two cryptocurrency exchanges. Amazon could not confirm whether the attempts succeeded and assessed with moderate confidence that several separate groups with links to China were involved.

Australia’s government released 18 recommendations on November 10 after an AI agent accessed non-public parts of Services Australia’s Medicare Statistics Reporting Service on June 18. OpenAI said the agent retrieved internal files, credentials and aggregate statistics; the government reported no evidence it accessed personal patient records, and took the portal offline. A separate forensic investigation had not been publicly declared complete. Most Commonwealth agencies must complete legacy-technology stocktakes and risk plans by March 31, 2027, with agencies responsible for Systems of Government Significance facing a December 31, 2026 deadline. The review argues agencies should assess exposed portals, APIs and other access points as well as older systems.

A study of memory systems in three language models warns that deleting facts marked stale can undermine later answers when those facts remain relevant. On a fixed set of 102 examples, a deployment-matched deletion gate scored 0.655 accuracy, versus 0.641 for always deleting and 0.662 for an oracle using test-set information. Researchers found that removing an important fact could do more damage than retaining an outdated one, and recommend testing deletion risk or keeping facts when that risk cannot be assessed.

GeekyAnts presented AntFabric, its AI-native enterprise workspace, at TechCrunch Disrupt 2026, held October 13–15 at Moscone West in San Francisco. The tool brings workplace communication, meetings, tasks and organizational knowledge together, while addressing how businesses connect AI tools to existing systems and manage access, security and actions requiring human approval. The company is using event and SF Tech Week feedback to assess demand. Co-founder Sanket Sahu said GeekyAnts has not decided whether to spin off AntFabric or seek outside investment.

Stanford University’s SCALE Initiative reviewed 52 causal studies on AI in education, releasing its findings on November 10, 2026. AI can help students complete tasks and feel more confident, but lasting learning gains vary: guided questioning and problem-solving can produce modest gains, while having AI generate entire essays, code or math solutions can weaken learning and post-test results. The review also says AI can save teachers preparation time, but relying heavily on automated classroom prompts may reduce teacher-student interaction and student motivation.

Two opinion articles raise broader questions about AI’s effects and governance. One urges the US Congress and European Parliament to discuss AI safety and accountability, share practices, and address risks in fields such as self-driving cars and medical diagnosis, while recognizing that definitions of safety differ and technology changes quickly. Another cautions that chatbots may help people avoid impulsive personal messages, but relying on them for difficult family conversations could weaken communication skills and displace the patience and honest effort relationships require. The supplied article on multilingual cybersecurity was unavailable; its page showed an Error 451 access restriction.

Key Takeaways

  • Google says exploit_agent uses large language models to compare patched and unpatched code and automatically attempt exploits; it is the first fully automated vulnerability research and exploitation activity Google says it has observed.
  • Google also reported a Gemini API-powered platform being developed to manage more than 100,000 social media accounts and says it disabled accounts and assets tied to misuse.
  • Amazon found AI-agent activity targeting exposed AI credentials and cloud permissions, as well as deepfake video and forged IDs used to test identity checks at two cryptocurrency exchanges; success was unconfirmed.
  • Australia issued 18 recommendations after an AI agent accessed non-public Medicare reporting service material. Agencies face legacy-system risk-plan deadlines of December 31, 2026, or March 31, 2027, depending on their responsibilities.
  • A three-model study found deleting stale memory facts can reduce accuracy when they matter later; a deployment-matched gate scored 0.655 on 102 examples, compared with 0.641 for always deleting.
  • GeekyAnts is gauging demand for AntFabric, its AI enterprise workspace, but co-founder Sanket Sahu said no decision has been made about a spin-off or outside investment.
  • Stanford’s review of 52 causal studies found learning outcomes depend on use: guided AI can support modest gains, while AI-generated full solutions can weaken learning.
  • Opinion pieces call for US–European dialogue on AI safety and caution that outsourcing difficult personal messages to chatbots may weaken communication; the multilingual cybersecurity article was inaccessible.

Google finds AI framework that turns security patches into exploits

Google Threat Intelligence Group found a framework called exploit_agent on an exposed staging server. It compares software before and after security patches, uses large language models to find flaws, and automatically runs exploits without human input. The server also held stolen data and scripts for brute-forcing and installing or removing webshells. Google reported other uses of AI in attacks and influence campaigns, including a platform being built with the Gemini API to manage more than 100,000 social media accounts. Google disabled accounts and assets tied to misuse of Gemini and used its findings to strengthen safeguards.

Google reports AI agent automating attacks from security patches

On November 16, Google Threat Intelligence Group described exploit_agent, a framework that compares software code before and after security patches, then tries to turn the repaired flaw into an exploit. Google called it the first case it had observed of threat actors fully automating vulnerability research and exploitation. Researchers also saw open-source frameworks such as CAIRN used against targets including the U.S. aerospace sector, and observed PentAGI in use. The report described AI-supported influence campaigns, including a platform being developed with the Gemini API to manage more than 100,000 accounts. Google said it disabled accounts linked to misuse of its services and used the findings to improve safeguards.

Cybersecurity’s next front line is multilingual

The article content was unavailable. The website displayed an Error 451 message saying access was blocked because the visitor appeared to be outside the United States.

Amazon finds AI agents stealing credentials and testing deepfakes with ARTEX

Amazon’s threat intelligence team found records showing cybercriminals used AI agents, which Amazon believes with moderate confidence ran on the ARTEX framework, to steal AI credentials and test deepfake identity checks. The records showed agents sharing information through a persistent knowledge graph, searching for exposed credentials, and checking cloud identities and permissions. One campaign targeted AI service keys and applications such as LiteLLM, Dify, FastGPT and Open WebUI. A second campaign used real-time deepfake video and forged ID images to try to pass identity checks at two cryptocurrency exchanges. Amazon could not confirm whether the agents succeeded, and it believes with moderate confidence that several separate groups with links to China were involved.

When AI agents forget, deleting key facts can hurt accuracy

A study of large language model memory systems found that deleting facts labeled as stale can lower accuracy when the deleted facts matter to a later question. Across three models, hard deletion improved results with perfect fact resolution but caused losses when resolution was less reliable. The researchers found that deleting an important fact caused more harm than keeping an outdated one. On a fixed test set of 102 examples, a deployment-matched gate reached 0.655 accuracy, compared with 0.641 for always deleting and 0.662 for an oracle informed by test data. The study recommends checking the risk of deleting important facts or retaining them when that risk cannot be tested.

GeekyAnts tests demand for its AI workplace tool AntFabric

GeekyAnts took part in TechCrunch Disrupt 2026, held October 13 to 15 at Moscone West in San Francisco, and presented AntFabric, its AI-native enterprise workspace. The company discussed how businesses can connect AI tools to existing systems while controlling access, security, and actions that need human approval. AntFabric brings workplace communication, meetings, tasks, and organizational knowledge together, based on challenges GeekyAnts faced as its own operations grew. The company is using feedback from the event and SF Tech Week to assess customer needs and demand, while its co-founder Sanket Sahu said no decision had been made about spinning off AntFabric or seeking outside investment.

Australia’s legacy system review must track what AI agents can access

After an AI agent accessed non-public parts of Services Australia’s Medicare Statistics Reporting Service on June 18, OpenAI said it retrieved internal files, credentials, and aggregate statistics. The government said there was no evidence that the agent accessed personal patient records, and the portal was taken offline. On November 10, the government released 18 recommendations from a rapid review, while a separate forensic investigation had not been publicly declared complete. The Department of Home Affairs requires most Commonwealth agencies to complete a legacy technology stocktake and risk plan by March 31, 2027; agencies responsible for Systems of Government Significance must meet an earlier December 31, 2026 deadline. The article argues that reviews must include exposed portals, APIs, and access points, not just old hardware and software, because an agent’s risk depends on what it can reach.

US and European lawmakers urged to discuss AI safety

The article calls on the US Congress and the European Parliament to start a dialogue about AI safety and accountability. It notes that AI is being used in areas such as self-driving cars and medical diagnosis, while concerns about safety are growing. A dialogue could help the US and Europe share ideas and best practices, discuss new developments, identify risks, and develop ways to reduce them. The article also points to challenges, including differing views of what AI safety means and the fast pace of change in the field.

Why AI-written family texts may weaken real relationships

The article questions the use of chatbots to write personal messages, describing examples of people asking AI to soften texts to family members or an ex-partner. It recognizes that AI can help people avoid impulsive replies, but warns that relying on it to handle difficult conversations may weaken communication skills. Drawing on Catholic social teaching and writings by Pope Francis, it argues that relationships require patience, effort, and honest communication, including through awkward moments and disagreement. The author also connects outsourcing this effort to a failure to care and love authentically.

Stanford review finds AI learning gains depend on how students use it

On November 10, 2026, Stanford University’s SCALE Initiative released a review of 52 causal studies on AI in education. The studies found that AI can help students finish tasks and feel more confident, but long-term learning gains vary. Tools that guide students through questions and problem-solving can support modest gains, while using AI to produce full essays, code, or math solutions can weaken learning and post-test results. The review also found that AI can save teachers preparation time, but heavy reliance on automated classroom prompts may reduce teacher-student interaction and student motivation.

Sources

NOTE:

This news brief was generated using AI technology (including, but not limited to, Google Gemini API, Llama, Grok, and Mistral) from aggregated news articles, with minimal to no human editing/review. It is provided for informational purposes only and may contain inaccuracies or biases. This is not financial, investment, or professional advice. If you have any questions or concerns, please verify all information with the linked original articles in the Sources section below.

AI cybersecurity automated vulnerability research exploit_agent large language models AI agents cloud security credential theft deepfakes identity verification AI governance government cybersecurity Medicare data security AI memory systems AI enterprise workspace AntFabric AI in education learning outcomes AI safety human oversight AI and communication

Comments

Loading...