## Security-Privacy-Cryptography

### 密码学在计算机领域的应用

#### hash function
##### 1.性质

- Deterministic: the same input always generates the same output.
- Non-invertible: it is hard to find an input `m` such that `hash(m) = h` for some desired output `h`.
- Target collision resistant: given an input `m_1`, it’s hard to find a different input `m_2` such that `hash(m_1) = hash(m_2)`.
- Collision resistant: it’s hard to find two inputs `m_1` and `m_2` such that `hash(m_1) = hash(m_2)` (note that this is a strictly stronger property than target collision resistance).

注意区分Target collision resistant和Collision resistant


##### 2.原理

```c++
int sum=0; 
for(int i=0; i<v.size(); ++i) sum=sum*131+v[i]; 
return sum;
```

* 对`vector<int>`做hash的方法：

  I. 用上面的方法，选取质数131，可能需要再设另一个质数取模

  II. 两个或三个的简单情形，可以利用pair和map

  III. 对于每一个整数， 把0\~7、8\~15、 16\~23、 24\~31的位置取出来变成char，cat之后再hash

* 方法I中取模用质数更好的原因

  * “ 如果p是一个质数，n是任意非零整数（不是p的倍数）， 那么px+ny=z对于任意的x,y,z都有解”， 这样可以保证取模相对均匀一些， 避免所谓的 primary clustering， 要证明这个需要引理：“方程 ax+by=1 有整数解当且仅当 a 和 b 互质”
* 哈希算法可能用到乘除法。模素数的剩余系除去 0 ，这个集合关于乘法构成群。只有群才能保证每个元素都有逆元，除法才能合法。假设要计算 (p / q) mod m，如果想让结果与 (p mod m) / (q mod m) 相等，必须令 m 为素数，否则逆元求不出来。

##### 3.应用
* Git中的id是由SHA-1 hash生成，40个16进制字符
  * SHA-1: 160bit 
  * SHA-2: 有不同位数，比如SHA-256
  * `$ printf 'hello' | sha1sum`

* [Commitment scheme](https://en.wikipedia.org/wiki/Commitment_scheme)



### Security and Cryptography

> [6.NULL Security and Cryptography](https://missing.csail.mit.edu/2020/security/)

[Cryptographic Right Answers](https://latacora.micro.blog/2018/04/03/cryptographic-right-answers.html)

#### Intro

##### Entropy
online guessing - 40 bits of entropy

offline guessing - 80 bits of entropy

##### Hash functions

 [lifetimes of cryptographic hash functions](https://valerieaurora.org/hash.html) 

##### Key derivation functions(KDFs)

应用：

- Producing keys from passphrases for use in other cryptographic algorithms (e.g. symmetric cryptography, see below).
- Storing login credentials. Storing plaintext passwords is bad; the right approach is to generate and store a random [salt](https://en.wikipedia.org/wiki/Salt_(cryptography)) `salt = random()` for each user, store `KDF(password + salt)`, and verify login attempts by re-computing the KDF given the entered password and the stored salt.

#### Symmetric cryptography

应用：

- Encrypting files for storage in an untrusted cloud service. This can be combined with KDFs, so you can encrypt a file with a passphrase. Generate `key = KDF(passphrase)`, and then store `encrypt(file, key)`.

```sh
openssl aes-256-cbc -salt -in {input filename} -out {output filename}
openssl aes-256-cbc -d -in {input filename} -out {output filename}
```



#### Asymmetric cryptography

用private key来sign，用public key来encrypt

- [PGP email encryption](https://en.wikipedia.org/wiki/Pretty_Good_Privacy). People can have their public keys posted online (e.g. in a PGP keyserver, or on [Keybase](https://keybase.io/)). Anyone can send them encrypted email.
- Private messaging. Apps like [Signal](https://signal.org/) and [Keybase](https://keybase.io/) use asymmetric keys to establish private communication channels.
- Signing software. Git can have GPG-signed commits and tags. With a posted public key, anyone can verify the authenticity of downloaded software.

* Key distribution: Asymmetric-key cryptography is wonderful, but it has a big challenge of distributing public keys / mapping public keys to real-world identities. There are many solutions to this problem. Signal has one simple solution: trust on first use, and support out-of-band public key exchange (you verify your friends’ “safety numbers” in person). PGP has a different solution, which is [web of trust](https://en.wikipedia.org/wiki/Web_of_trust). Keybase has yet another solution of [social proof](https://keybase.io/blog/chat-apps-softer-than-tofu) (along with other neat ideas). Each model has its merits; we (the instructors) like Keybase’s model.

应用：

* In use, once the server knows the client’s public key (stored in the `.ssh/authorized_keys` file), a connecting client can prove its identity using asymmetric signatures. This is done through [challenge-response](https://en.wikipedia.org/wiki/Challenge–response_authentication). At a high level, the server picks a random number and sends it to the client. The client then signs this message and sends the signature back to the server, which checks the signature against the public key on record. This effectively proves that the client is in possession of the private key corresponding to the public key that’s in the server’s `.ssh/authorized_keys` file, so the server can allow the client to log in.
* [Set up GPG](https://www.digitalocean.com/community/tutorials/how-to-use-gpg-to-encrypt-and-sign-messages)

[在Github上使用GPG的全过程 - 林溪的文章 - 知乎](https://zhuanlan.zhihu.com/p/76861431)

```sh
sudo apt-get install gnupg
gpg --gen-key
gpg --output ~/revocation.crt --gen-revoke your_email@address.com
chmod 600 ~/revocation.crt

gpg --import name_of_pub_key_file
https://pgp.mit.edu/

gpg --keyserver pgp.mit.edu  --search-keys search_parameters
gpg --fingerprint your_email@address.com

# 信任（用自己的密钥为其签名验证）
gpg --sign-key email@example.com
gpg --output ~/signed.key --export --armor email@example.com
gpg --import ~/signed.key

gpg --output ~/mygpg.key --armor --export your_email@address.com

gpg --keyserver pgp.mit.edu  --send-keys ...
gpg --keyserver pgp.mit.edu  --recv-keys ...

gpg --encrypt --sign --armor -r person@email.com name_of_file
# 如果想自己decrypt，需要第二个-r recipient
gpg file_name.asc

gpg --list-keys
gpg --refresh-keys

gpg --keyserver key_server --refresh-keys

```

Git and GPG
```sh
git config --global user.signingkey {key_id}
git config --global commit.gpgsign true

git commit -S
git tag -s

git log/show --show-signature
git tag -v
```

信任Github所用的GPG密钥，使本地确认在Github网页端进行的操作的真实性
```sh
curl https://github.com/web-flow.gpg | gpg --import
gpg --sign-key 4AEE18F83AFDEB23
```

### Security and Privacy

> [6.NULL Security and Privacy](https://missing.csail.mit.edu/2019/security/)

#### Intro

Follow the [right people](https://heimdalsecurity.com/blog/best-twitter-cybersec-accounts/)

使用安全的密码管理器，比如[1password](https://1password.com/), [KeePass](https://keepass.info/), [KeePass](https://keepass.info/), [`pass`](https://www.passwordstore.org/)

更安全的two-factor authentication双因素认证：a [FIDO/U2F](https://fidoalliance.org/) dongle (a [YubiKey](https://www.yubico.com/quiz/) for example, which has [20% off for students](https://www.yubico.com/why-yubico/for-education/)). TOTP (like Google Authenticator or Duo) will also work in a pinch, but [doesn’t protect against phishing](https://twitter.com/taviso/status/1082015009348104192). SMS is pretty much useless unless your threat model only includes random strangers picking up your password in transit. 

* [SMS's issue](https://www.kaspersky.com/blog/2fa-practical-guide/24219/)

##### General Security Advice

Tech Solidarity has a pretty great list of [do’s and don’ts for journalists](https://techsolidarity.org/resources/basic_security.htm) that has a lot of sane advice, and is decently up-to-date. @thegrugq also has a good blog post on [travel security advice](https://medium.com/@thegrugq/stop-fabricating-travel-security-advice-35259bf0e869) that’s worth reading. We’ll repeat much of the advice from those sources here, plus some more. Also, get a [USB data blocker](https://amzn.com/B00QRRZ2QM), because [USB is scary](https://www.bleepingcomputer.com/news/security/heres-a-list-of-29-different-types-of-usb-attacks/).



#### Private Communication

Use [Signal](https://www.signal.org/) ([setup instructions](https://medium.com/@mshelton/signal-for-beginners-c6b44f76a1f0). [Wire](https://wire.com/en/) is [fine too](https://www.securemessagingapps.com/); WhatsApp is okay; [don’t use Telegram](https://twitter.com/bascule/status/897187286554628096) (不错的文章)). Desktop messengers are pretty broken (partially due to usually relying on Electron, which is a huge trust stack).

E-mail is particularly problematic, even if PGP signed. It’s not generally forward-secure, and the key-distribution problem is pretty severe. [keybase.io](https://keybase.io/) helps, and is useful for a number of other reasons. Also, PGP keys are generally handled on desktop computers, which is one of the least secure computing environments. Relatedly, consider getting a Chromebook, or just work on a tablet with a keyboard.



#### File Security

File security is hard, and operates on many level. What is it you’re trying to secure against?

[![$5 wrench](Security-Privacy-Cryptography/security.png)](https://xkcd.com/538/)

- Offline attacks (someone steals your laptop while it’s off): turn on full disk encryption. ([cryptsetup + LUKS](https://wiki.archlinux.org/index.php/Dm-crypt/Encrypting_a_non-root_file_system) on Linux, [BitLocker](https://fossbytes.com/enable-full-disk-encryption-windows-10/) on Windows, [FileVault](https://support.apple.com/en-us/HT204837) on macOS. Note that this won’t help if the attacker *also* has you and really wants your secrets.

- Online attacks (someone has your laptop and it’s on): use file encryption. There are two primary mechanisms for doing so    

  - Encrypted filesystems: stacked filesystem encryption software  encrypts files individually rather than having encrypted block devices.  You can “mount” these filesystems by providing the decryption key, and  then browse the files inside it freely. When you unmount it, those files are all unavailable.  Modern solutions include [gocryptfs](https://github.com/rfjakob/gocryptfs) and [eCryptFS](http://ecryptfs.org/). More detailed comparisons can be found [here](https://nuetzlich.net/gocryptfs/comparison/) and [here](https://wiki.archlinux.org/index.php/disk_encryption#Comparison_table)
  - Encrypted files: encrypt individual files with symmetric encryption (see `gpg -c`) and a secret key. Or, like `pass`, also encrypt the key with your public key so only you can read it back later with your private key. Exact encryption settings matter a lot!

- [Plausible deniability](https://en.wikipedia.org/wiki/Plausible_deniability) (what seems to be the problem officer?): usually lower performance, and easier to lose data. Hard to actually prove that it provides [deniable encryption](https://en.wikipedia.org/wiki/Deniable_encryption)! See the [discussion here](https://security.stackexchange.com/questions/135846/is-plausible-deniability-actually-feasible-for-encrypted-volumes-disks), and then consider whether you may want to try [VeraCrypt](https://www.veracrypt.fr/en/Home.html) (the maintained fork of good ol’ TrueCrypt).

- Encrypted backups: use  [Tarsnap](https://www.tarsnap.com/) or [Borgbase](https://www.borgbase.com/)

  - Think about whether an attacker can delete your backups if they get a hold of your laptop!

#### 进程安全

##### 最小权限原则

**核心原则**: 任何程序、任何用户都只应拥有其完成任务所必需的最小权限 (Principle of Least Privilege)。以 root 用户运行服务是极大的安全风险。

*   **应用服务 (Tomcat, Redis, Kafka)**: **绝对不应该**以 root 身份运行。
    *   **原因**: 这些服务完全不需要 root 权限。如果服务本身或其上运行的应用被攻破，攻击者将直接获得整个服务器的 root 权限。
    *   **特例**: 如果服务需要监听特权端口（< 1024，如80端口），可以 root *启动*，但必须在端口绑定后立即**降权**，将工作进程切换到低权限用户（如 `www-data`）运行。

*   **系统/安全服务 (HostGuard)**: **通常必须**以 root 身份运行。
    *   **原因**: 其核心功能（如监控所有进程、管理防火墙、扫描文件系统）要求必须具备系统级的最高权限才能有效执行。

#### Internet Security & Privacy

The internet is a *very* scary place. Open WiFi networks [are](https://www.troyhunt.com/the-beginners-guide-to-breaking-website/) [scary](https://www.troyhunt.com/talking-with-scott-hanselman-on/). Make sure you delete them afterwards, otherwise your phone will happily announce and re-connect to something with the same name later!

If you’re ever on a network you don’t trust, a VPN *may* be worthwhile, but keep in mind that you’re trusting the VPN provider *a lot*. Do you really trust them more than your ISP? If you truly want a VPN, use a provider you’re sure you trust, and you should probably pay for it. Or set up [WireGuard](https://www.wireguard.com/) for yourself – it’s [excellent](https://latacora.micro.blog/there-will-be/)!

If you’re particularly privacy-oriented, [privacytools.io](https://privacytools.io) is also a good resource.

Some of you may wonder about [Tor](https://www.torproject.org/). Keep in mind that Tor is *not* particularly resistant to powerful global attackers, and is weak against traffic analysis attacks. It may be useful for hiding traffic on a small scale, but won’t really buy you all that much in terms of privacy. You’re better off using more secure services in the first place (Signal, TLS + certificate pinning, etc.).


##### 常见高危端口

在网络安全中，某些端口因其关联的服务非常核心或存在固有弱点，而成为攻击者的重点扫描和攻击目标。

*   **21 (FTP - 文件传输协议)**: 主要风险在于其默认以**明文传输**数据和用户凭证，极易被网络嗅探工具截获。
*   **22 (SSH - 安全外壳协议)**: 协议本身安全，但作为服务器远程管理的主要入口，是**暴力破解攻击**的常见目标。安全策略包括：禁用密码登录（改用密钥）、禁止root直接登录、更改默认端口。
*   **3389 (RDP - 远程桌面协议)**: Windows远程桌面服务的默认端口。因其广泛使用和通常具备高权限，是勒索软件和黑客攻击的重点目标。
*   **3306 (MySQL)**: MySQL数据库服务的默认端口。直接暴露在公网是极大的安全隐患，容易导致数据泄露或被攻击。


#### Web Security

> 架构层补充来源：[Cloudflare Project Glasswing](https://blog.cloudflare.com/cyber-frontier-models/)。漏洞防御不能只优化 patch 速度；更重要的是让 bug 存在时也难以被利用：在应用前用输入验证、WAF / protocol guard 阻断可达路径；用最小权限与组件隔离限制单点缺陷的横向影响；用统一 rollout 让修复同时抵达所有运行实例。披露窗口的风险更接近 `external reachability × blast radius × rollout inconsistency`，而不只是 `time-to-patch`。完整 agent 漏洞发现与验证漏斗见 [Cloudflare Vulnerability Harness](./AI-Applied-Algorithms.md#cloudflare-vulnerability-harness从-security-skill-到跨仓库控制面)。

So, you want to go on the Web too? Jeez, you’re really pushing your luck here.

Install [HTTPS Everywhere](https://www.eff.org/https-everywhere). SSL/TLS is [critical](https://www.troyhunt.com/ssl-is-not-about-encryption/) (已读, **Login Landing Page Must Use SSL**), and it’s *not* just about encryption, but also about being able to verify that you’re talking to the right service in the first place! If you run your own web server, [test it](https://www.ssllabs.com/ssltest/index.html). TLS configuration [can get hairy](https://wiki.mozilla.org/Security/Server_Side_TLS). HTTPS Everywhere will do its very best to never navigate you to HTTP sites when there’s an alternative. That doesn’t save you, but it helps. If you’re truly paranoid, blacklist any SSL/TLS CAs that you don’t absolutely need.

Install [uBlock Origin](https://github.com/gorhill/uBlock). It is a [wide-spectrum blocker](https://github.com/gorhill/uBlock/wiki/Blocking-mode) that doesn’t just stop ads, but all sorts of third-party communication a page may try to do. And inline scripts and such. If you’re willing to spend some time on configuration to make things work, go to [medium mode](https://github.com/gorhill/uBlock/wiki/Blocking-mode:-medium-mode) or even [hard mode](https://github.com/gorhill/uBlock/wiki/Blocking-mode:-hard-mode). Those *will* make some sites not work until you’ve fiddled with the settings enough, but will also significantly improve your online security.

If you’re using Firefox, enable [Multi-Account Containers](https://support.mozilla.org/en-US/kb/containers). Create separate containers for social networks, banking, shopping, etc. Firefox will keep the cookies and other state for each of the containers totally separate, so sites you visit in one container can’t snoop on sensitive data from the others. In Google Chrome, you can use [Chrome Profiles](https://support.google.com/chrome/answer/2364824) to achieve similar results.


### AI 服务供应链与 Agent 滥用

来源：[Anthropic 威胁情报报告，2026 年 9 月](https://www.anthropic.com/threat-intelligence-report-september-2026)。报告选取 2025 年 12 月至 2026 年 8 月发现并处置的突出案例，属于服务商调查材料，不代表典型使用情况；组织归因与数据流转指控未经本文独立核实。

**AI 放大的既有能力，也有执行规模。** 工具调用、并行 agent、跨会话记忆和反馈重试，让少数操作者能持续处理多个陌生环境。复杂工作流不再能单独证明操作者技术高超；评价增益要分别看速度、规模和能力深度。Agent 评测还需区分请求、生成、执行与现实结果，见 [Agent 评估与安全](./AI-Applied-Algorithms.md#agent-评估与安全)。

#### 供应链既是入口，也是资产

AI API key 和会话令牌有三重价值：转卖获利、使用受害者的计算额度、借合法账户掩盖来源。模型代理、客户端、评测沙箱及其生产凭证因此都进入攻击面。

报告中的代表案例：

- **评测沙箱泄密（GTG-50020）**：Anthropic 称，恶意指令使某 AI 厂商的自动评测沙箱交出其持有的生产 API 密钥。关键边界是“不可信任务内容能否触达真实凭证”，而不仅是模型能否识别恶意请求。
- **伪装低价 AI 服务（GTG-50021）**：报告描述了冒充折扣模型服务、替换实际模型并窃取用户凭证的活动。价格和界面名称都不能证明后端身份与客户端可信性。
- **虚假约会应用（GTG-15001）**：报告称某应用网络使用大量 AI 身份与用户交谈，却宣传真人服务。风险来自规模化运营和身份欺骗，不要求模型先获得突破性能力。

工程原则：生产凭证留在受信执行层，模型侧只得到受限工具或不透明句柄；沙箱按任务授予最小权限，限制网络出口、预算与敏感数据访问。模型拒绝、环境隔离和异常检测分别承担不同职责，不能互相替代。

#### 模型名称不等于数据处理边界

**工程师内部系统案例（GTG-16002）**：Anthropic 称，一名工程师使用 Kimi 开发企业内部系统时提交了内部代码和多家企业的有效访问凭证，相关请求被转发到 Claude。`live credentials` 表示当时仍可用于认证的凭证；公开材料没有说明具体权限，也没有证明这些凭证随后被用于入侵。

这个案例涉及两道独立授权：**允许把数据提交给某服务，不等于允许该服务再转交其他处理方。** 报告对用户是否获告知存在不确定表述，因此不能把所有客户均未获告知、所有请求均被转发写成已证实事实。

调用链可拆为：

```text
用户 → 应用 / 模型路由 → 名义供应商 → 实际推理供应商 → 返回结果
                         └→ 交互保存 / 数据整理 / 训练用途（需分别核验）
```

推理转发、交互留存、用于训练是三种不同处理行为，不能仅凭发生第一种就认定后两种。数据治理应核对实际处理方、路由与 fallback、保存期限、训练用途和删除机制；凭证应在请求、工具参数、日志与附件入口统一脱敏。

#### 反蒸馏与证据边界

蒸馏本身是正常训练方法；报告针对的是未经授权的能力提取及其伴随的账户、凭证和数据滥用。Preserved thinking 通过约束受保护推理之前的上下文改写，缩小诱导模型重新输出推理轨迹的路径；最终答案仍可能提供训练信号，不能据此声称阻止所有蒸馏。协议机制见 [推理记录与上下文绑定](./AI-Applied-Algorithms.md#stealing-reasoning-traces-与-external-thinking推理记录是可提取的侧信道)。

读取威胁报告时保留三条证据纪律：**观察到请求不等于任务成功；使用 AI 不等于证明净增益；封禁账户不等于受害系统恢复。** 服务商案例能揭示具体风险路径，但缺少对照实验与总体分母时，不能据此计算全社会风险增长或模型的因果贡献。

#### 主动模型归因：用生成数字偏差反推后端身份（ModelTrace）

> 来源：[ModelTrace 浏览器本地版](https://xqy2006.github.io/ModelTrace/) 与 [xqy2006/ModelTrace](https://github.com/xqy2006/ModelTrace)（MIT，2026-09-15 读取，commit `3f0dd2f4`；含 `static/` 纯前端版与 Codex 插件 `codex-plugin/modeltrace-guard/`）；前置工作参考 [hlwy-ai-checker](https://github.com/hanlinwenyuan/hlwy-ai-checker)。整理时间：2026-09-17。

上一节 GTG-50021 的边界是「价格和界面名称都不能证明后端身份」。ModelTrace 把这条边界变成可操作的检测：**不问供应商要证据，而是让模型自己产出可统计的偏差**。做法是三条独立的「长整数生成」挑战（每次 218–333 个 `[1, 355]` 区间整数，要求逐项凭第一反应、禁止工具 / 计算器 / 搜索 / 计数递增 / 等差 / 循环 / 重复区块），把输出分布当成指纹，在候选库里做归因。

- **指纹表示**：`355` 维计数分布（取值区间 `[1, 355]`，α = 0.5 平滑）做 Hellinger 特征，叠加有序块特征（序列切 4 段各 16 桶 + 末位数字分布）。总分 `0.75 × Hellinger 模型中心相似度 + 0.25 × 有序块特征`（库内 `ordered_block_weight = 0.25`）。
- **环境去偏是核心设计**：建库时把各共享环境下特征的平均偏移做 SVD，取前 `2` 个方向当 nuisance basis；归因时先投影掉这些方向再与全部模型中心比较。这样「换了 system prompt / 中英文 / 上下文长度」造成的整体位移不会被误算成模型差异。挑战套件本身也刻意铺了 12 组环境（通道 `clean / user / system` × 格式 `json / 中文 / 英文` × 前缀 `96 / 512 / 2048` 词）。
- **概率与校准**：三份回答分别打分取平均，再用与查询数配套的温度做全局 softmax（库内 β：1 次查询 `6.84`、2 次 `12.0`、3 次 `12.0`；对应分组交叉验证准确率 `0.955 / 0.996 / 1.000`）。家族概率是该家族具体模型概率之和；单次检测要求至少 80 个有效数字。
- **闭集边界**：结果是「当前候选库内、均匀先验」的闭集概率，**库外模型照样会被归到最相似的候选**，这是最常见的误用点。当前库 13 个模型（GPT 6：`gpt-5.4 / 5.5 / 5.6-sol / -terra / -luna / gpt-6-astra`；Claude 7：`haiku-4-5 / sonnet-4-6 / sonnet-5 / opus-4-6 / -4-7 / -4-8 / opus-5`）。GPT 采自官方订阅 Codex，Claude 采自第三方中转 OAIPro——**Claude 侧指纹刻画的是那条渠道的行为，不等于官方 API 特征**。
- **作者自述的失效条件**：结果仅供参考、不是决定性证据；system prompt 会强烈影响数字偏好，因此在 Claude Code 里测得的结果偏差较大，官方建议不要在 Claude Code 内测试。API 自动测试支持 OpenAI Chat Completions 与 Anthropic Messages 两种格式，Key 只用于当次请求、不落盘。

对工程判断的意义：

- **模型身份是需要主动验证的运行时属性，而不是配置项**：路由、fallback、渠道替换、量化版本差异都会改变实际后端，而客户端能拿到的只有行为，所以「行为统计 + 参考库」是少数可落地的证据形式。
- **它的证据强度由可探测能力决定**：只有三条挑战、闭集候选库、且依赖环境去偏，因此适合当 canary / 可疑信号，不适合当判定结论；把它当开放集识别使用会系统性高估把握度。
- 与上面 [模型名称不等于数据处理边界](#模型名称不等于数据处理边界) 是同一问题的两面：那一节说「名义供应商 ≠ 实际处理方」，这一节给的是「用输出行为反推实际处理方」的一个具体手段。
- 同一指纹库还被封装成任务内监测（按工具调用间隔 fork 快照做后台复测），用于发现**任务进行中后端模型被静默替换**，机制见 [AI-Agent-Engineering.md - ModelTrace Guard](./AI-Agent-Engineering.md#modeltrace-guard把后端模型被静默替换做成任务内可探测信号)。

### Cryptography I, Stanford University, Dan Boneh

* [coursera课程](https://www.coursera.org/learn/crypto/home/welcome)
* [密码学资源推荐](https://blog.cryptographyengineering.com/useful-cryptography-resources/)
* [A Graduate Course In Applied Cryptography](https://toc.cryptobook.us/)
* [A Computational Introduction to Number Theory and Algebra](https://www.shoup.net/ntb/)



### Potpourri

* [Python沙盒逃逸](https://ciphersaw.me/ctf-wiki/pwn/linux/sandbox/python-sandbox-escape/)
  * `__builtins__`

* SSRF (Server-Side Request Forgery) attack
  * 本质上，不允许访问内网资源即可修复，但可能会被301/302/307/308重定向、DNS重绑定攻破，修复的时候有可能手抖。
  * 方案：可以用安全开发包，从传输层彻底断掉内网请求，提供自定义黑白名单功能

* RCE (Remote Code Execution) attach
  * 根本原因：开发者使用python的eval进行json解析，而本身eval是用来执行一个字符串表达式，并返回表达式的值，这意味着可以执行任何python代码，从而执行系统命令。
  * 方案：通过AST Node类型识别，干掉不安全的执行

* 任意文件读取/下载漏洞
* SQL注入漏洞
  * 编码不规范引发。攻击者拼接SQL片段，通过返回包内容的大小，逐步获取数据库的内容
* [炮打Ollama鉴权机制——为什么CS从业者都需要学习网络安全](https://www.bilibili.com/video/BV1eZ421z79W/)